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/unaligned.h>
16 #include <xnnpack/vadd.h>
17
18
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_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_vadd_minmax_ukernel__avx_mul32_ld32_x32(
20 size_t n,
21 const int8_t* input_a,
22 const int8_t* input_b,
23 int8_t* output,
24 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse4_mul32.bias);
27 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
28 const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.b_multiplier);
29 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse4_mul32.shift);
30 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
31 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
32 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
33
34 for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
35 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
36 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
37 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
38 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
39 const __m128i va89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 8)));
40 const __m128i vb89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 8)));
41 const __m128i vaCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 12)));
42 const __m128i vbCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 12)));
43 const __m128i vaGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 16)));
44 const __m128i vbGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 16)));
45 const __m128i vaKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 20)));
46 const __m128i vbKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 20)));
47 const __m128i vaOPQR = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 24)));
48 const __m128i vbOPQR = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 24)));
49 const __m128i vaSTUV = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 28)));
50 const __m128i vbSTUV = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 28)));
51 input_a += 32;
52 input_b += 32;
53
54 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
55 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
56 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_mullo_epi32(va89AB, va_multiplier));
57 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_mullo_epi32(vaCDEF, va_multiplier));
58 __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_mullo_epi32(vaGHIJ, va_multiplier));
59 __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_mullo_epi32(vaKLMN, va_multiplier));
60 __m128i vaccOPQR = _mm_add_epi32(vbias, _mm_mullo_epi32(vaOPQR, va_multiplier));
61 __m128i vaccSTUV = _mm_add_epi32(vbias, _mm_mullo_epi32(vaSTUV, va_multiplier));
62
63 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
64 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
65 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vb89AB, vb_multiplier));
66 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vbCDEF, vb_multiplier));
67 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vbGHIJ, vb_multiplier));
68 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vbKLMN, vb_multiplier));
69 vaccOPQR = _mm_add_epi32(vaccOPQR, _mm_mullo_epi32(vbOPQR, vb_multiplier));
70 vaccSTUV = _mm_add_epi32(vaccSTUV, _mm_mullo_epi32(vbSTUV, vb_multiplier));
71
72 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
73 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
74 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
75 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
76 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
77 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
78 vaccOPQR = _mm_sra_epi32(vaccOPQR, vshift);
79 vaccSTUV = _mm_sra_epi32(vaccSTUV, vshift);
80
81 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
82 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
83 const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
84 const __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
85
86 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
87 __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
88
89 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
90 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
91
92 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
93 voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
94
95 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
96 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
97 output += 32;
98 }
99 if XNN_UNLIKELY(n != 0) {
100 do {
101 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
102 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
103 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
104 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
105 input_a += 8;
106 input_b += 8;
107
108 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
109 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
110
111 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
112 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
113
114 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
115 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
116
117 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
118
119 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
120 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
121 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
122
123 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
124 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
125 output += 8;
126 n -= 8 * sizeof(int8_t);
127 } else {
128 if (n & (4 * sizeof(int8_t))) {
129 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
130 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
131 output += 4;
132 }
133 if (n & (2 * sizeof(int8_t))) {
134 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
135 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
136 output += 2;
137 }
138 if (n & (1 * sizeof(int8_t))) {
139 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
140 }
141 n = 0;
142 }
143 } while (n != 0);
144 }
145 }
146