1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmulc/sse-mul16-ld64.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <smmintrin.h>
13
14 #include <xnnpack/vmul.h>
15
16
xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8(
18 size_t n,
19 const uint8_t* input_a,
20 const uint8_t* input_b,
21 uint8_t* output,
22 const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23
24 {
25 const __m128i va_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.a_zero_point);
26 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
27 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
28 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
29 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max);
30
31 __m128i vxb = _mm_sub_epi16(
32 _mm_shuffle_epi32(_mm_cvtsi32_si128(UINT32_C(0x00010001) * (uint32_t) (uint16_t) (int16_t) *input_b), 0),
33 _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point));
34 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
35 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
36 input_a += 8;
37
38
39 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
40
41 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
42 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);
43
44 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
45 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
46
47 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
48 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
49
50 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
51 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
52
53 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
54 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
55
56 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
57
58
59 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
60
61 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
62
63 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
64
65 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
66 output += 8;
67 }
68 if XNN_UNLIKELY(n != 0) {
69 {
70 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
71
72
73 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
74
75 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
76 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);
77
78 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
79 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
80
81 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
82 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
83
84 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
85 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
86
87 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
88 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
89
90 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
91
92 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
93 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
94 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
95
96 if (n & (4 * sizeof(uint8_t))) {
97 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
98 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
99 output += 4;
100 }
101 if (n & (2 * sizeof(uint8_t))) {
102 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
103 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
104 output += 2;
105 }
106 if (n & (1 * sizeof(uint8_t))) {
107 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
108 }
109 }
110 }
111 }
112