Debian Codegraph @

src/qu8-vcvt/gen/qu8-vcvt-sse2-u32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vcvt/sse2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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 <tmmintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/vcvt.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qu8_vcvt_ukernel__sse2_u32(
20 size_t batch,
21 const uint8_t* input,
22 uint8_t* output,
23 const struct xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 const __m128i vmultiplier = _mm_set1_epi16(params->scalar.multiplier);
31 const __m128i vbias = _mm_set1_epi32(
32 (int32_t) (((uint32_t) (int32_t) params->scalar.output_zero_point) << 8) -
33 (int32_t) params->scalar.multiplier * (int32_t) params->scalar.input_zero_point +
34 INT32_C(0x80));
35 XNN_FORCE_REALIZATION(vmultiplier);
36 XNN_FORCE_REALIZATION(vbias);
37 const __m128i vzero = _mm_setzero_si128();
38 for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) {
39 const __m128i vx0 = _mm_loadu_si128((const __m128i*) input);
40 const __m128i vx1 = _mm_loadu_si128((const __m128i*) (input + 16));
41 input += 32;
42
43 const __m128i vextx0 = _mm_unpacklo_epi8(vx0, vzero);
44 const __m128i vextx1 = _mm_unpackhi_epi8(vx0, vzero);
45 const __m128i vextx2 = _mm_unpacklo_epi8(vx1, vzero);
46 const __m128i vextx3 = _mm_unpackhi_epi8(vx1, vzero);
47
48 const __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier);
49 const __m128i vprodhi0 = _mm_mulhi_epu16(vextx0, vmultiplier);
50 const __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier);
51 const __m128i vprodhi1 = _mm_mulhi_epu16(vextx1, vmultiplier);
52 const __m128i vprodlo2 = _mm_mullo_epi16(vextx2, vmultiplier);
53 const __m128i vprodhi2 = _mm_mulhi_epu16(vextx2, vmultiplier);
54 const __m128i vprodlo3 = _mm_mullo_epi16(vextx3, vmultiplier);
55 const __m128i vprodhi3 = _mm_mulhi_epu16(vextx3, vmultiplier);
56
57 __m128i vacc0 = _mm_unpacklo_epi16(vprodlo0, vprodhi0);
58 __m128i vacc1 = _mm_unpackhi_epi16(vprodlo0, vprodhi0);
59 __m128i vacc2 = _mm_unpacklo_epi16(vprodlo1, vprodhi1);
60 __m128i vacc3 = _mm_unpackhi_epi16(vprodlo1, vprodhi1);
61 __m128i vacc4 = _mm_unpacklo_epi16(vprodlo2, vprodhi2);
62 __m128i vacc5 = _mm_unpackhi_epi16(vprodlo2, vprodhi2);
63 __m128i vacc6 = _mm_unpacklo_epi16(vprodlo3, vprodhi3);
64 __m128i vacc7 = _mm_unpackhi_epi16(vprodlo3, vprodhi3);
65
66 vacc0 = _mm_add_epi32(vacc0, vbias);
67 vacc1 = _mm_add_epi32(vacc1, vbias);
68 vacc2 = _mm_add_epi32(vacc2, vbias);
69 vacc3 = _mm_add_epi32(vacc3, vbias);
70 vacc4 = _mm_add_epi32(vacc4, vbias);
71 vacc5 = _mm_add_epi32(vacc5, vbias);
72 vacc6 = _mm_add_epi32(vacc6, vbias);
73 vacc7 = _mm_add_epi32(vacc7, vbias);
74
75 vacc0 = _mm_srai_epi32(vacc0, 8);
76 vacc1 = _mm_srai_epi32(vacc1, 8);
77 vacc2 = _mm_srai_epi32(vacc2, 8);
78 vacc3 = _mm_srai_epi32(vacc3, 8);
79 vacc4 = _mm_srai_epi32(vacc4, 8);
80 vacc5 = _mm_srai_epi32(vacc5, 8);
81 vacc6 = _mm_srai_epi32(vacc6, 8);
82 vacc7 = _mm_srai_epi32(vacc7, 8);
83
84 vacc0 = _mm_packs_epi32(vacc0, vacc1);
85 vacc1 = _mm_packs_epi32(vacc2, vacc3);
86 vacc2 = _mm_packs_epi32(vacc4, vacc5);
87 vacc3 = _mm_packs_epi32(vacc6, vacc7);
88
89 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
90 const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
91
92 _mm_storeu_si128((__m128i*) output, vy0);
93 _mm_storeu_si128((__m128i*) (output + 16), vy1);
94 output += 32;
95 }
96 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
97 const __m128i vx = _mm_loadu_si128((const __m128i*) input);
98 input += 16;
99
100 const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vzero);
101 const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vzero);
102
103 const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier);
104 const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier);
105 const __m128i vprodhi_lo = _mm_mulhi_epu16(vextx_lo, vmultiplier);
106 const __m128i vprodhi_hi = _mm_mulhi_epu16(vextx_hi, vmultiplier);
107
108 __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo);
109 __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo);
110 __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi);
111 __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi);
112
113 vacc_ll = _mm_add_epi32(vacc_ll, vbias);
114 vacc_lh = _mm_add_epi32(vacc_lh, vbias);
115 vacc_hl = _mm_add_epi32(vacc_hl, vbias);
116 vacc_hh = _mm_add_epi32(vacc_hh, vbias);
117
118 vacc_ll = _mm_srai_epi32(vacc_ll, 8);
119 vacc_lh = _mm_srai_epi32(vacc_lh, 8);
120 vacc_hl = _mm_srai_epi32(vacc_hl, 8);
121 vacc_hh = _mm_srai_epi32(vacc_hh, 8);
122
123 const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh);
124 const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh);
125
126 const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
127 _mm_storeu_si128((__m128i*) output, vy);
128 output += 16;
129 }
130 if XNN_UNLIKELY(batch != 0) {
131 assert(batch >= 1 * sizeof(uint8_t));
132 assert(batch <= 15 * sizeof(uint8_t));
133
134 const __m128i vx = _mm_loadu_si128((const __m128i*) input);
135
136 const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vzero);
137 const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vzero);
138
139 const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier);
140 const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier);
141 const __m128i vprodhi_lo = _mm_mulhi_epu16(vextx_lo, vmultiplier);
142 const __m128i vprodhi_hi = _mm_mulhi_epu16(vextx_hi, vmultiplier);
143
144 __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo);
145 __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo);
146 __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi);
147 __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi);
148
149 vacc_ll = _mm_add_epi32(vacc_ll, vbias);
150 vacc_lh = _mm_add_epi32(vacc_lh, vbias);
151 vacc_hl = _mm_add_epi32(vacc_hl, vbias);
152 vacc_hh = _mm_add_epi32(vacc_hh, vbias);
153
154 vacc_ll = _mm_srai_epi32(vacc_ll, 8);
155 vacc_lh = _mm_srai_epi32(vacc_lh, 8);
156 vacc_hl = _mm_srai_epi32(vacc_hl, 8);
157 vacc_hh = _mm_srai_epi32(vacc_hh, 8);
158
159 const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh);
160 const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh);
161
162 __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
163 if (batch & (8 * sizeof(uint8_t))) {
164 _mm_storel_epi64((__m128i*) output, vy);
165 vy = _mm_unpackhi_epi64(vy, vy);
166 output += 8;
167 }
168 if (batch & (4 * sizeof(uint8_t))) {
169 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vy));
170 vy = _mm_srli_epi64(vy, 32);
171 output += 4;
172 }
173 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
174 if (batch & (2 * sizeof(uint8_t))) {
175 unaligned_store_u16(output, (uint16_t) vy_lo);
176 vy_lo >>= 16;
177 output += 2;
178 }
179 if (batch & (1 * sizeof(uint8_t))) {
180 *output = (uint8_t) vy_lo;
181 }
182 }
183 }