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_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)])18 void xnn_qs8_vadd_minmax_ukernel__avx_mul32_ld32_x32(
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 >= 32 * sizeof(int8_t); n -= 32 * 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 const __m128i vaGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 16));
43 const __m128i vbGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 16));
44 const __m128i vaKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 20));
45 const __m128i vbKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 20));
46 const __m128i vaOPQR = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 24));
47 const __m128i vbOPQR = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 24));
48 const __m128i vaSTUV = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 28));
49 const __m128i vbSTUV = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 28));
50 input_a += 32;
51 input_b += 32;
52
53 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
54 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
55 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_mullo_epi32(va89AB, va_multiplier));
56 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_mullo_epi32(vaCDEF, va_multiplier));
57 __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_mullo_epi32(vaGHIJ, va_multiplier));
58 __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_mullo_epi32(vaKLMN, va_multiplier));
59 __m128i vaccOPQR = _mm_add_epi32(vbias, _mm_mullo_epi32(vaOPQR, va_multiplier));
60 __m128i vaccSTUV = _mm_add_epi32(vbias, _mm_mullo_epi32(vaSTUV, va_multiplier));
61
62 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
63 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
64 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vb89AB, vb_multiplier));
65 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vbCDEF, vb_multiplier));
66 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vbGHIJ, vb_multiplier));
67 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vbKLMN, vb_multiplier));
68 vaccOPQR = _mm_add_epi32(vaccOPQR, _mm_mullo_epi32(vbOPQR, vb_multiplier));
69 vaccSTUV = _mm_add_epi32(vaccSTUV, _mm_mullo_epi32(vbSTUV, vb_multiplier));
70
71 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
72 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
73 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
74 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
75 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
76 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
77 vaccOPQR = _mm_sra_epi32(vaccOPQR, vshift);
78 vaccSTUV = _mm_sra_epi32(vaccSTUV, vshift);
79
80 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
81 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
82 const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
83 const __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
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 va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
101 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b));
102 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
103 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 4));
104 input_a += 8;
105 input_b += 8;
106
107 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
108 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
109
110 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
111 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
112
113 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
114 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
115
116 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
117
118 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
119 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
120 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
121
122 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
123 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
124 output += 8;
125 n -= 8 * sizeof(int8_t);
126 } else {
127 if (n & (4 * sizeof(int8_t))) {
128 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
129 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
130 output += 4;
131 }
132 if (n & (2 * sizeof(int8_t))) {
133 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
134 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
135 output += 2;
136 }
137 if (n & (1 * sizeof(int8_t))) {
138 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
139 }
140 n = 0;
141 }
142 } while (n != 0);
143 }
144 }
145