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_x16(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_x16(
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 >= 16 * sizeof(int8_t); n -= 16 * 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 input_a += 16;
40 input_b += 16;
41
42 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
43 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
44 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_mullo_epi32(va89AB, va_multiplier));
45 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_mullo_epi32(vaCDEF, va_multiplier));
46
47 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
48 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
49 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
50 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
51
52 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
53 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
54
55 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
56
57 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
58
59 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
60
61 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
62 output += 16;
63 }
64 if XNN_UNLIKELY(n != 0) {
65 do {
66 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
67 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
68 input_a += 8;
69
70 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
71 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
72
73 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
74 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
75
76 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
77
78 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
79 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
80 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
81
82 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
83 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
84 output += 8;
85 n -= 8 * sizeof(int8_t);
86 } else {
87 if (n & (4 * sizeof(int8_t))) {
88 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
89 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
90 output += 4;
91 }
92 if (n & (2 * sizeof(int8_t))) {
93 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
94 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
95 output += 2;
96 }
97 if (n & (1 * sizeof(int8_t))) {
98 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
99 }
100 n = 0;
101 }
102 } while (n != 0);
103 }
104 }
105