1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmulc/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/vmul.h>
15
16
xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_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_vmulc_minmax_fp32_ukernel__sse41_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 __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
27 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
28 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
29 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max);
30
31 __m128i vxb = _mm_sub_epi16(
32 _mm_shuffle_epi32(_mm_cvtsi32_si128(UINT32_C(0x00010001) * (uint32_t) (uint16_t) (int16_t) *input_b), 0),
33 _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point));
34 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
35 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
36 const __m128i va89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
37 input_a += 16;
38
39
40 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
41 const __m128i vxa89ABCDEF = _mm_sub_epi16(va89ABCDEF, va_zero_point);
42
43 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
44 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);
45 const __m128i vprod89ABCDEFlo = _mm_mullo_epi16(vxa89ABCDEF, vxb);
46 const __m128i vprod89ABCDEFhi = _mm_mulhi_epi16(vxa89ABCDEF, vxb);
47
48 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
49 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
50 const __m128i vprod89AB = _mm_unpacklo_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
51 const __m128i vprodCDEF = _mm_unpackhi_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
52
53 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
54 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
55 __m128 vfpacc89AB = _mm_cvtepi32_ps(vprod89AB);
56 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vprodCDEF);
57
58 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
59 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
60 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
61 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
62
63 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
64 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
65 const __m128i vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
66 const __m128i vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
67
68 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
69 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
70
71
72 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
73
74 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
75
76 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
77
78 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
79 output += 16;
80 }
81 if XNN_UNLIKELY(n != 0) {
82 do {
83 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
84 input_a += 8;
85
86
87 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
88
89 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
90 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);
91
92 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
93 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
94
95 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
96 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
97
98 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
99 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
100
101 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
102 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
103
104 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
105
106 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
107 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
108 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
109
110 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
111 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
112 output += 8;
113 n -= 8 * sizeof(uint8_t);
114 } else {
115 if (n & (4 * sizeof(uint8_t))) {
116 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
117 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
118 output += 4;
119 }
120 if (n & (2 * sizeof(uint8_t))) {
121 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
122 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
123 output += 2;
124 }
125 if (n & (1 * sizeof(uint8_t))) {
126 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
127 }
128 n = 0;
129 }
130 } while (n != 0);
131 }
132 }
133