Debian Codegraph @

src/qs8-vmulc/gen/qs8-vmulc-minmax-fp32-sse2-mul16-ld64-u8.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 <emmintrin.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__sse2_mul16_ld64_u8(
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_epi16(params->scalar.output_min);
36 const __m128i voutput_max = _mm_set1_epi16(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 >= 8 * sizeof(int8_t); batch -= 8 * sizeof(int8_t)) {
47 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
48 input_a += 8;
49
50 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
51
52 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
53
54 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
55 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);
56
57 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
58 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
59
60 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
61 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
62
63 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
64 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
65
66 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
67 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
68
69 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
70
71 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
72
73 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
74
75 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
76
77
78 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
79 output += 8;
80 }
81 if XNN_UNLIKELY(batch != 0) {
82 {
83 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
84
85 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
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 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
106 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
107
108 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
109
110 if (batch & (4 * sizeof(int8_t))) {
111 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
112 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
113 output += 4;
114 }
115 if (batch & (2 * sizeof(int8_t))) {
116 unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
117 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
118 output += 2;
119 }
120 if (batch & (1 * sizeof(int8_t))) {
121 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
122 }
123 }
124 }
125 }