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_x24(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_x24(
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 >= 24 * sizeof(int8_t); n -= 24 * 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 const __m128i va89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 8)));
44 const __m128i vaCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 12)));
45 const __m128i vaGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 16)));
46 const __m128i vaKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 20)));
47 input_a += 24;
48 input_b += 24;
49
50 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
51 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
52 __m128i vacc89AB = _mm_macc_epi32(va89AB, va_multiplier, vbias);
53 __m128i vaccCDEF = _mm_macc_epi32(vaCDEF, va_multiplier, vbias);
54 __m128i vaccGHIJ = _mm_macc_epi32(vaGHIJ, va_multiplier, vbias);
55 __m128i vaccKLMN = _mm_macc_epi32(vaKLMN, va_multiplier, vbias);
56
57 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
58 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
59 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
60 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
61 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
62 vaccKLMN = _mm_sra_epi32(vaccKLMN, 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
68 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
69 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
70
71 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
72 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
73
74 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
75 voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
76
77 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
78 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
79 output += 24;
80 }
81 if XNN_UNLIKELY(n != 0) {
82 do {
83 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
84 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
85 input_a += 8;
86
87 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
88 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
89
90 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
91 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
92
93 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
94
95 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
96 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
97 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
98
99 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
100 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
101 output += 8;
102 n -= 8 * sizeof(int8_t);
103 } else {
104 if (n & (4 * sizeof(int8_t))) {
105 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
106 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
107 output += 4;
108 }
109 if (n & (2 * sizeof(int8_t))) {
110 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
111 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
112 output += 2;
113 }
114 if (n & (1 * sizeof(int8_t))) {
115 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
116 }
117 n = 0;
118 }
119 } while (n != 0);
120 }
121 }
122