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/vmul.h>
15
16
xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16(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__sse2_mul16_ld64_x16(
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 >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
33 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
34 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
35 __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
36 __m128i vb89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_b + 8));
37 input_a += 16;
38 input_b += 16;
39
40 const __m128i vzero = _mm_setzero_si128();
41 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
42 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
43 va89ABCDEF = _mm_unpacklo_epi8(va89ABCDEF, vzero);
44 vb89ABCDEF = _mm_unpacklo_epi8(vb89ABCDEF, vzero);
45
46 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
47 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
48 const __m128i vxa89ABCDEF = _mm_sub_epi16(va89ABCDEF, va_zero_point);
49 const __m128i vxb89ABCDEF = _mm_sub_epi16(vb89ABCDEF, vb_zero_point);
50
51 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
52 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
53 const __m128i vprod89ABCDEFlo = _mm_mullo_epi16(vxa89ABCDEF, vxb89ABCDEF);
54 const __m128i vprod89ABCDEFhi = _mm_mulhi_epi16(vxa89ABCDEF, vxb89ABCDEF);
55
56 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
57 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
58 const __m128i vprod89AB = _mm_unpacklo_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
59 const __m128i vprodCDEF = _mm_unpackhi_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
60
61 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
62 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
63 __m128 vfpacc89AB = _mm_cvtepi32_ps(vprod89AB);
64 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vprodCDEF);
65
66 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
67 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
68 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
69 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
70
71 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
72 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
73 const __m128i vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
74 const __m128i vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
75
76 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
77 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
78
79
80 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
81
82 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
83
84 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
85
86 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
87 output += 16;
88 }
89 if XNN_UNLIKELY(n != 0) {
90 do {
91 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
92 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
93 input_a += 8;
94 input_b += 8;
95
96 const __m128i vzero = _mm_setzero_si128();
97 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
98 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
99
100 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
101 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
102
103 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
104 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
105
106 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
107 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
108
109 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
110 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
111
112 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
113 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
114
115 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
116 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
117
118 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
119
120 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
121 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
122 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
123
124 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
125 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
126 output += 8;
127 n -= 8 * sizeof(uint8_t);
128 } else {
129 if (n & (4 * sizeof(uint8_t))) {
130 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
131 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
132 output += 4;
133 }
134 if (n & (2 * sizeof(uint8_t))) {
135 *((uint16_t*) output) = (uint16_t) _mm_cvtsi128_si32(vout0123456701234567);
136 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
137 output += 2;
138 }
139 if (n & (1 * sizeof(uint8_t))) {
140 *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
141 }
142 n = 0;
143 }
144 } while (n != 0);
145 }
146 }
147