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 #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_vaddc_minmax_ukernel__xop_mul32_ld32_x8(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_vaddc_minmax_ukernel__xop_mul32_ld32_x8(
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 va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
32 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse4_mul32.shift);
33 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
34 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
35 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
36
37 __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
38 vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
39 vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
40 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
41 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
42 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
43 input_a += 8;
44 input_b += 8;
45
46 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
47 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
48
49 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
50 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
51
52 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
53
54 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
55
56 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
57
58 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
59
60 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
61 output += 8;
62 }
63 if XNN_UNLIKELY(n != 0) {
64 {
65 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
66 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
67
68 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
69 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
70
71 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
72 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
73
74 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
75
76 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
77 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
78 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
79
80 if (n & (4 * sizeof(int8_t))) {
81 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
82 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
83 output += 4;
84 }
85 if (n & (2 * sizeof(int8_t))) {
86 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
87 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
88 output += 2;
89 }
90 if (n & (1 * sizeof(int8_t))) {
91 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
92 }
93 }
94 }
95 }
96