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/unaligned.h>
15 #include <xnnpack/vmul.h>
16
17
xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8(
19 size_t n,
20 const int8_t* input_a,
21 const int8_t* input_b,
22 int8_t* output,
23 const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24
25 {
26 const __m128i va_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.a_zero_point);
27 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.b_zero_point);
28 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
29 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
30 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
31 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse4.output_max);
32
33 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
34 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
35 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
36 input_a += 8;
37 input_b += 8;
38
39
40 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
41 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
42
43 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
44 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
45
46 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
47 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
48
49 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
50 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
51
52 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
53 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
54
55 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
56 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
57
58 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
59
60
61 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
62
63 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
64
65 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
66
67 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
68 output += 8;
69 }
70 if XNN_UNLIKELY(n != 0) {
71 {
72 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
73 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
74
75
76 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
77 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
78
79 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
80 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
81
82 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
83 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
84
85 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
86 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
87
88 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
89 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
90
91 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
92 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
93
94 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
95
96 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
97 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
98 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
99
100 if (n & (4 * sizeof(int8_t))) {
101 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
102 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
103 output += 4;
104 }
105 if (n & (2 * sizeof(int8_t))) {
106 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
107 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
108 output += 2;
109 }
110 if (n & (1 * sizeof(int8_t))) {
111 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
112 }
113 }
114 }
115 }
116