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__sse41_mul32_ld32_x8(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__sse41_mul32_ld32_x8(
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 >= 8 * sizeof(int8_t); n -= 8 * 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 input_a += 8;
38 input_b += 8;
39
40 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
41 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
42
43 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
44 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
45
46 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
47
48 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
49
50 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
51
52 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
53
54 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
55 output += 8;
56 }
57 if XNN_UNLIKELY(n != 0) {
58 {
59 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
60 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
61
62 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
63 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
64
65 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
66 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
67
68 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
69
70 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
71 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
72 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
73
74 if (n & (4 * sizeof(int8_t))) {
75 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
76 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
77 output += 4;
78 }
79 if (n & (2 * sizeof(int8_t))) {
80 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
81 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
82 output += 2;
83 }
84 if (n & (1 * sizeof(int8_t))) {
85 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
86 }
87 }
88 }
89 }
90