1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmul/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_vmul_minmax_fp32_ukernel__avx_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_vmul_minmax_fp32_ukernel__avx_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 __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point);
27 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
28 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
29 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
30 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max);
31
32 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
33 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
34 const __m128i vb01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
35 input_a += 8;
36 input_b += 8;
37
38
39 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
40 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
41
42 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
43 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
44
45 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
46 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
47
48 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
49 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
50
51 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
52 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
53
54 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
55 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
56
57 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
58
59
60 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
61
62 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
63
64 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
65
66 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
67 output += 8;
68 }
69 if XNN_UNLIKELY(n != 0) {
70 {
71 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
72 const __m128i vb01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
73
74
75 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
76 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
77
78 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
79 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
80
81 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
82 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
83
84 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
85 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
86
87 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
88 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
89
90 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
91 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
92
93 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
94
95 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
96 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
97 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
98
99 if (n & (4 * sizeof(uint8_t))) {
100 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
101 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
102 output += 4;
103 }
104 if (n & (2 * sizeof(uint8_t))) {
105 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
106 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
107 output += 2;
108 }
109 if (n & (1 * sizeof(uint8_t))) {
110 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
111 }
112 }
113 }
114 }
115