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_qu8_vadd_minmax_ukernel__avx_mul32_ld32_x8(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_vadd_minmax_ukernel__avx_mul32_ld32_x8(
19 size_t n,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const union xnn_qu8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse4.bias);
26 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4.a_multiplier);
27 const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse4.b_multiplier);
28 const __m128i vshift = _mm_loadu_si32(params->sse4.shift);
29 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
30 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
31 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4.output_max);
32
33 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
34 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(input_a));
35 const __m128i vb0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(input_b));
36 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(input_a + 4));
37 const __m128i vb4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(input_b + 4));
38 input_a += 8;
39 input_b += 8;
40
41 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
42 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
43
44 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
45 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
46
47 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
48 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
49
50 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
51
52 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
53
54 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
55
56 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
57
58 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
59 output += 8;
60 }
61 if XNN_UNLIKELY(n != 0) {
62 {
63 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(input_a));
64 const __m128i vb0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(input_b));
65 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(input_a + 4));
66 const __m128i vb4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(input_b + 4));
67
68 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
69 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
70
71 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
72 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
73
74 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
75 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
76
77 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
78
79 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
80 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
81 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
82
83 if (n & (4 * sizeof(uint8_t))) {
84 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
85 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
86 output += 4;
87 }
88 if (n & (2 * sizeof(uint8_t))) {
89 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
90 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
91 output += 2;
92 }
93 if (n & (1 * sizeof(uint8_t))) {
94 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
95 }
96 }
97 }
98 }
99