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/unaligned.h>
16 #include <xnnpack/vadd.h>
17
18
xnn_qu8_vaddc_minmax_ukernel__sse41_mul32_ld32_x8(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_vaddc_minmax_ukernel__sse41_mul32_ld32_x8(
20 size_t n,
21 const uint8_t* input_a,
22 const uint8_t* input_b,
23 uint8_t* output,
24 const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4.a_multiplier);
27 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse4.shift);
28 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
29 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
30 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4.output_max);
31
32 __m128i vbias = _mm_cvtsi32_si128(params->sse4.b_multiplier[0] * (int32_t) *input_b);
33 vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
34 vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4.bias));
35 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
36 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
37 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 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_sra_epi32(vacc0123, vshift);
45 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
46
47 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
48
49 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
50
51 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
52
53 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
54
55 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
56 output += 8;
57 }
58 if XNN_UNLIKELY(n != 0) {
59 {
60 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
61 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
62
63 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
64 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
65
66 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
67 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
68
69 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
70
71 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
72 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
73 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
74
75 if (n & (4 * sizeof(uint8_t))) {
76 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
77 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
78 output += 4;
79 }
80 if (n & (2 * sizeof(uint8_t))) {
81 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
82 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
83 output += 2;
84 }
85 if (n & (1 * sizeof(uint8_t))) {
86 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
87 }
88 }
89 }
90 }
91