Debian Codegraph @

src/qs8-vmulc/gen/qs8-vmulc-minmax-fp32-sse41-mul16-ld64-u16.c

raw
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/common.h"
15 #include "xnnpack/unaligned.h"
16 #include "xnnpack/vbinary.h"
17
18
19 void xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_u16(
20 size_t batch,
21 const int8_t* input_a,
22 const int8_t* input_b,
23 int8_t* output,
24 const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(batch != 0);
27 assert(batch % sizeof(int8_t) == 0);
28 assert(input_a != NULL);
29 assert(input_b != NULL);
30 assert(output != NULL);
31
32 const __m128i va_zero_point = _mm_set1_epi16(params->scalar.a_zero_point);
33 const __m128 vscale = _mm_set1_ps(params->scalar.scale);
34 const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point);
35 const __m128i voutput_min = _mm_set1_epi8(params->scalar.output_min);
36 const __m128i voutput_max = _mm_set1_epi8(params->scalar.output_max);
37
38 __m128i vxb = _mm_set1_epi16((UINT32_C(0x00010001) * (uint32_t) (uint16_t) (int16_t) *input_b) - params->scalar.b_zero_point);
39
40 XNN_FORCE_REALIZATION(va_zero_point);
41 XNN_FORCE_REALIZATION(vscale);
42 XNN_FORCE_REALIZATION(voutput_zero_point);
43 XNN_FORCE_REALIZATION(voutput_min);
44 XNN_FORCE_REALIZATION(voutput_max);
45
46 for (; batch >= 16 * sizeof(int8_t); batch -= 16 * sizeof(int8_t)) {
47 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
48 const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
49 input_a += 16;
50
51
52 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
53 const __m128i vxa89ABCDEF = _mm_sub_epi16(va89ABCDEF, va_zero_point);
54
55 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
56 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);
57 const __m128i vprod89ABCDEFlo = _mm_mullo_epi16(vxa89ABCDEF, vxb);
58 const __m128i vprod89ABCDEFhi = _mm_mulhi_epi16(vxa89ABCDEF, vxb);
59
60 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
61 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
62 const __m128i vprod89AB = _mm_unpacklo_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
63 const __m128i vprodCDEF = _mm_unpackhi_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
64
65 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
66 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
67 __m128 vfpacc89AB = _mm_cvtepi32_ps(vprod89AB);
68 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vprodCDEF);
69
70 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
71 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
72 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
73 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
74
75 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
76 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
77 const __m128i vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
78 const __m128i vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
79
80 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
81 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
82
83
84 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
85
86 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
87
88 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
89
90 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
91 output += 16;
92 }
93 if XNN_UNLIKELY(batch != 0) {
94 do {
95 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
96 input_a += 8;
97
98
99 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
100
101 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
102 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);
103
104 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
105 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
106
107 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
108 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
109
110 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
111 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
112
113 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
114 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
115
116 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
117
118 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
119 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
120 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
121
122 if XNN_LIKELY(batch >= (8 * sizeof(int8_t))) {
123 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
124 output += 8;
125 batch -= 8 * sizeof(int8_t);
126 } else {
127 if (batch & (4 * sizeof(int8_t))) {
128 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
129 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
130 output += 4;
131 }
132 if (batch & (2 * sizeof(int8_t))) {
133 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
134 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
135 output += 2;
136 }
137 if (batch & (1 * sizeof(int8_t))) {
138 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
139 }
140 batch = 0;
141 }
142 } while (batch != 0);
143 }
144 }