1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/sse-mul32-ld32.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 <immintrin.h>
13
14 #include <xnnpack/intrinsics-polyfill.h>
15 #include <xnnpack/vaddsub.h>
16
17
xnn_qs8_vadd_minmax_ukernel__avx_mul32_ld32_x16(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)])18 void xnn_qs8_vadd_minmax_ukernel__avx_mul32_ld32_x16(
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_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse4_mul32.bias);
26 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
27 const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.b_multiplier);
28 const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
29 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
30 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
31 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
32
33 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
34 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
35 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b));
36 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
37 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 4));
38 const __m128i va89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 8));
39 const __m128i vb89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 8));
40 const __m128i vaCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 12));
41 const __m128i vbCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 12));
42 input_a += 16;
43 input_b += 16;
44
45 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
46 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
47 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_mullo_epi32(va89AB, va_multiplier));
48 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_mullo_epi32(vaCDEF, va_multiplier));
49
50 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
51 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
52 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vb89AB, vb_multiplier));
53 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vbCDEF, vb_multiplier));
54
55 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
56 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
57 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
58 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
59
60 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
61 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
62
63 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
64
65 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
66
67 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
68
69 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
70 output += 16;
71 }
72 if XNN_UNLIKELY(n != 0) {
73 do {
74 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
75 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b));
76 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
77 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 4));
78 input_a += 8;
79 input_b += 8;
80
81 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
82 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
83
84 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
85 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
86
87 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
88 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
89
90 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
91
92 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
93 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
94 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
95
96 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
97 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
98 output += 8;
99 n -= 8 * sizeof(int8_t);
100 } else {
101 if (n & (4 * sizeof(int8_t))) {
102 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
103 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
104 output += 4;
105 }
106 if (n & (2 * sizeof(int8_t))) {
107 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
108 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
109 output += 2;
110 }
111 if (n & (1 * sizeof(int8_t))) {
112 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
113 }
114 n = 0;
115 }
116 } while (n != 0);
117 }
118 }
119