1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/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_vaddc_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_vaddc_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 va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
26 const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
27 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
28 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
29 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
30
31 __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
32 vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
33 vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
34 for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
35 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
36 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
37 const __m128i va89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 8));
38 const __m128i vaCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 12));
39 const __m128i vaGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 16));
40 const __m128i vaKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 20));
41 const __m128i vaOPQR = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 24));
42 const __m128i vaSTUV = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 28));
43 input_a += 32;
44 input_b += 32;
45
46 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
47 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
48 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_mullo_epi32(va89AB, va_multiplier));
49 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_mullo_epi32(vaCDEF, va_multiplier));
50 __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_mullo_epi32(vaGHIJ, va_multiplier));
51 __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_mullo_epi32(vaKLMN, va_multiplier));
52 __m128i vaccOPQR = _mm_add_epi32(vbias, _mm_mullo_epi32(vaOPQR, va_multiplier));
53 __m128i vaccSTUV = _mm_add_epi32(vbias, _mm_mullo_epi32(vaSTUV, va_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 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
60 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
61 vaccOPQR = _mm_sra_epi32(vaccOPQR, vshift);
62 vaccSTUV = _mm_sra_epi32(vaccSTUV, vshift);
63
64 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
65 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
66 const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
67 const __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
68
69 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
70 __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
71
72 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
73 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
74
75 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
76 voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
77
78 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
79 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
80 output += 32;
81 }
82 if XNN_UNLIKELY(n != 0) {
83 do {
84 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
85 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
86 input_a += 8;
87
88 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
89 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
90
91 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
92 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
93
94 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
95
96 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
97 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
98 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
99
100 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
101 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
102 output += 8;
103 n -= 8 * sizeof(int8_t);
104 } else {
105 if (n & (4 * sizeof(int8_t))) {
106 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
107 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
108 output += 4;
109 }
110 if (n & (2 * sizeof(int8_t))) {
111 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
112 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
113 output += 2;
114 }
115 if (n & (1 * sizeof(int8_t))) {
116 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
117 }
118 n = 0;
119 }
120 } while (n != 0);
121 }
122 }
123