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 <emmintrin.h>
13
14 #include <xnnpack/unaligned.h>
15 #include <xnnpack/vmul.h>
16
17
xnn_qs8_vmul_minmax_fp32_ukernel__sse2_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__sse2_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_sse2.a_zero_point);
27 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point);
28 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
29 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
30 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
31 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max);
32
33 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
34 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
35 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
36 input_a += 8;
37 input_b += 8;
38
39 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
40 vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
41
42 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
43 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
44
45 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
46 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
47
48 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
49 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
50
51 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
52 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
53
54 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
55 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
56
57 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
58 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
59
60 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
61
62 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
63
64 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
65
66 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
67
68
69 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
70 output += 8;
71 }
72 if XNN_UNLIKELY(n != 0) {
73 {
74 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
75 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
76
77 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
78 vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
79
80 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
81 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
82
83 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
84 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
85
86 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
87 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
88
89 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
90 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
91
92 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
93 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
94
95 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
96 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
97
98 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
99 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
100 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
101
102 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
103
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_cvtsi128_si32(vout0123456701234567));
111 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
112 output += 2;
113 }
114 if (n & (1 * sizeof(int8_t))) {
115 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
116 }
117 }
118 }
119 }
120