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_qu8_vmul_minmax_fp32_ukernel__sse2_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)])18 void xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8(
19 size_t n,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const union xnn_qu8_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(uint8_t); n -= 8 * sizeof(uint8_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 const __m128i vzero = _mm_setzero_si128();
40 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
41 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
42
43 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
44 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
45
46 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
47 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
48
49 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
50 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
51
52 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
53 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
54
55 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
56 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
57
58 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
59 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
60
61 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
62
63
64 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
65
66 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
67
68 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
69
70 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
71 output += 8;
72 }
73 if XNN_UNLIKELY(n != 0) {
74 {
75 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
76 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
77
78 const __m128i vzero = _mm_setzero_si128();
79 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
80 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
81
82 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
83 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
84
85 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
86 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
87
88 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
89 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
90
91 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
92 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
93
94 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
95 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
96
97 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
98 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
99
100 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
101
102 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
103 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
104 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
105
106 if (n & (4 * sizeof(uint8_t))) {
107 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
108 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
109 output += 4;
110 }
111 if (n & (2 * sizeof(uint8_t))) {
112 unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
113 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
114 output += 2;
115 }
116 if (n & (1 * sizeof(uint8_t))) {
117 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
118 }
119 }
120 }
121 }
122