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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/intrinsics-polyfill.h>
20 #include <xnnpack/unaligned.h>
21 #include <xnnpack/vadd.h>
22
23
xnn_qs8_vadd_minmax_ukernel__xop_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)])24 void xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32(
25 size_t n,
26 const int8_t* input_a,
27 const int8_t* input_b,
28 int8_t* output,
29 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse4_mul32.bias);
32 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
33 const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.b_multiplier);
34 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse4_mul32.shift);
35 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
36 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
37 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
38
39 for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
40 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
41 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
42 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
43 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
44 const __m128i va89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 8)));
45 const __m128i vb89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 8)));
46 const __m128i vaCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 12)));
47 const __m128i vbCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 12)));
48 const __m128i vaGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 16)));
49 const __m128i vbGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 16)));
50 const __m128i vaKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 20)));
51 const __m128i vbKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 20)));
52 const __m128i vaOPQR = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 24)));
53 const __m128i vbOPQR = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 24)));
54 const __m128i vaSTUV = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 28)));
55 const __m128i vbSTUV = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 28)));
56 input_a += 32;
57 input_b += 32;
58
59 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
60 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
61 __m128i vacc89AB = _mm_macc_epi32(va89AB, va_multiplier, vbias);
62 __m128i vaccCDEF = _mm_macc_epi32(vaCDEF, va_multiplier, vbias);
63 __m128i vaccGHIJ = _mm_macc_epi32(vaGHIJ, va_multiplier, vbias);
64 __m128i vaccKLMN = _mm_macc_epi32(vaKLMN, va_multiplier, vbias);
65 __m128i vaccOPQR = _mm_macc_epi32(vaOPQR, va_multiplier, vbias);
66 __m128i vaccSTUV = _mm_macc_epi32(vaSTUV, va_multiplier, vbias);
67
68 vacc0123 = _mm_macc_epi32(vb0123, vb_multiplier, vacc0123);
69 vacc4567 = _mm_macc_epi32(vb4567, vb_multiplier, vacc4567);
70 vacc89AB = _mm_macc_epi32(vb89AB, vb_multiplier, vacc89AB);
71 vaccCDEF = _mm_macc_epi32(vbCDEF, vb_multiplier, vaccCDEF);
72 vaccGHIJ = _mm_macc_epi32(vbGHIJ, vb_multiplier, vaccGHIJ);
73 vaccKLMN = _mm_macc_epi32(vbKLMN, vb_multiplier, vaccKLMN);
74 vaccOPQR = _mm_macc_epi32(vbOPQR, vb_multiplier, vaccOPQR);
75 vaccSTUV = _mm_macc_epi32(vbSTUV, vb_multiplier, vaccSTUV);
76
77 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
78 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
79 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
80 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
81 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
82 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
83 vaccOPQR = _mm_sra_epi32(vaccOPQR, vshift);
84 vaccSTUV = _mm_sra_epi32(vaccSTUV, vshift);
85
86 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
87 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
88 const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
89 const __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
90
91 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
92 __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
93
94 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
95 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
96
97 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
98 voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
99
100 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
101 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
102 output += 32;
103 }
104 if XNN_UNLIKELY(n != 0) {
105 do {
106 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
107 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
108 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
109 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
110 input_a += 8;
111 input_b += 8;
112
113 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
114 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
115
116 vacc0123 = _mm_macc_epi32(vb0123, vb_multiplier, vacc0123);
117 vacc4567 = _mm_macc_epi32(vb4567, vb_multiplier, vacc4567);
118
119 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
120 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
121
122 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
123
124 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
125 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
126 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
127
128 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
129 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
130 output += 8;
131 n -= 8 * sizeof(int8_t);
132 } else {
133 if (n & (4 * sizeof(int8_t))) {
134 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
135 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
136 output += 4;
137 }
138 if (n & (2 * sizeof(int8_t))) {
139 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
140 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
141 output += 2;
142 }
143 if (n & (1 * sizeof(int8_t))) {
144 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
145 }
146 n = 0;
147 }
148 } while (n != 0);
149 }
150 }
151