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_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)])18 void xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16(
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 >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
34 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
35 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
36 __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
37 __m128i vb89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_b + 8));
38 input_a += 16;
39 input_b += 16;
40
41 const __m128i vzero = _mm_setzero_si128();
42 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
43 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
44 va89ABCDEF = _mm_unpacklo_epi8(va89ABCDEF, vzero);
45 vb89ABCDEF = _mm_unpacklo_epi8(vb89ABCDEF, vzero);
46
47 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
48 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
49 const __m128i vxa89ABCDEF = _mm_sub_epi16(va89ABCDEF, va_zero_point);
50 const __m128i vxb89ABCDEF = _mm_sub_epi16(vb89ABCDEF, vb_zero_point);
51
52 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
53 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
54 const __m128i vprod89ABCDEFlo = _mm_mullo_epi16(vxa89ABCDEF, vxb89ABCDEF);
55 const __m128i vprod89ABCDEFhi = _mm_mulhi_epi16(vxa89ABCDEF, vxb89ABCDEF);
56
57 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
58 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
59 const __m128i vprod89AB = _mm_unpacklo_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
60 const __m128i vprodCDEF = _mm_unpackhi_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
61
62 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
63 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
64 __m128 vfpacc89AB = _mm_cvtepi32_ps(vprod89AB);
65 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vprodCDEF);
66
67 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
68 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
69 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
70 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
71
72 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
73 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
74 const __m128i vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
75 const __m128i vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
76
77 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
78 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
79
80
81 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
82
83 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
84
85 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
86
87 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
88 output += 16;
89 }
90 if XNN_UNLIKELY(n != 0) {
91 do {
92 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
93 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
94 input_a += 8;
95 input_b += 8;
96
97 const __m128i vzero = _mm_setzero_si128();
98 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
99 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
100
101 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
102 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
103
104 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
105 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
106
107 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
108 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
109
110 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
111 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
112
113 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
114 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
115
116 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
117 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
118
119 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
120
121 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
122 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
123 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
124
125 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
126 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
127 output += 8;
128 n -= 8 * sizeof(uint8_t);
129 } else {
130 if (n & (4 * sizeof(uint8_t))) {
131 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
132 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
133 output += 4;
134 }
135 if (n & (2 * sizeof(uint8_t))) {
136 unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
137 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
138 output += 2;
139 }
140 if (n & (1 * sizeof(uint8_t))) {
141 *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
142 }
143 n = 0;
144 }
145 } while (n != 0);
146 }
147 }
148