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_x16(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_x16(
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 >= 16 * sizeof(int8_t); n -= 16 * 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 const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
37 const __m128i vb89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_b + 8)));
38 input_a += 16;
39 input_b += 16;
40
41
42 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
43 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
44 const __m128i vxa89ABCDEF = _mm_sub_epi16(va89ABCDEF, va_zero_point);
45 const __m128i vxb89ABCDEF = _mm_sub_epi16(vb89ABCDEF, vb_zero_point);
46
47 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
48 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
49 const __m128i vprod89ABCDEFlo = _mm_mullo_epi16(vxa89ABCDEF, vxb89ABCDEF);
50 const __m128i vprod89ABCDEFhi = _mm_mulhi_epi16(vxa89ABCDEF, vxb89ABCDEF);
51
52 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
53 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
54 const __m128i vprod89AB = _mm_unpacklo_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
55 const __m128i vprodCDEF = _mm_unpackhi_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
56
57 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
58 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
59 __m128 vfpacc89AB = _mm_cvtepi32_ps(vprod89AB);
60 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vprodCDEF);
61
62 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
63 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
64 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
65 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
66
67 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
68 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
69 const __m128i vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
70 const __m128i vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
71
72 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
73 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
74
75
76 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
77
78 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
79
80 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
81
82 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
83 output += 16;
84 }
85 if XNN_UNLIKELY(n != 0) {
86 do {
87 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
88 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
89 input_a += 8;
90 input_b += 8;
91
92
93 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
94 const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point);
95
96 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);
97 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567);
98
99 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
100 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
101
102 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
103 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
104
105 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
106 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
107
108 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
109 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
110
111 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
112
113 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
114 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
115 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
116
117 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
118 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
119 output += 8;
120 n -= 8 * sizeof(int8_t);
121 } else {
122 if (n & (4 * sizeof(int8_t))) {
123 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
124 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
125 output += 4;
126 }
127 if (n & (2 * sizeof(int8_t))) {
128 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
129 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
130 output += 2;
131 }
132 if (n & (1 * sizeof(int8_t))) {
133 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
134 }
135 n = 0;
136 }
137 } while (n != 0);
138 }
139 }
140