1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/qs8-vcvt/sse4.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2022 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker #include <immintrin.h>
13*4bdc9457SAndroid Build Coastguard Worker
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/intrinsics-polyfill.h>
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vcvt.h>
17*4bdc9457SAndroid Build Coastguard Worker
18*4bdc9457SAndroid Build Coastguard Worker
xnn_qu8_vcvt_ukernel__sse41_x32(size_t n,const uint8_t * x,uint8_t * y,const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19*4bdc9457SAndroid Build Coastguard Worker void xnn_qu8_vcvt_ukernel__sse41_x32(
20*4bdc9457SAndroid Build Coastguard Worker size_t n,
21*4bdc9457SAndroid Build Coastguard Worker const uint8_t* x,
22*4bdc9457SAndroid Build Coastguard Worker uint8_t* y,
23*4bdc9457SAndroid Build Coastguard Worker const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24*4bdc9457SAndroid Build Coastguard Worker {
25*4bdc9457SAndroid Build Coastguard Worker assert(n != 0);
26*4bdc9457SAndroid Build Coastguard Worker assert(n % sizeof(uint8_t) == 0);
27*4bdc9457SAndroid Build Coastguard Worker assert(x != NULL);
28*4bdc9457SAndroid Build Coastguard Worker assert(y != NULL);
29*4bdc9457SAndroid Build Coastguard Worker
30*4bdc9457SAndroid Build Coastguard Worker const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.input_zero_point);
31*4bdc9457SAndroid Build Coastguard Worker const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->ssse3.multiplier);
32*4bdc9457SAndroid Build Coastguard Worker const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.output_zero_point);
33*4bdc9457SAndroid Build Coastguard Worker for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
34*4bdc9457SAndroid Build Coastguard Worker __m128i vacc0 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
35*4bdc9457SAndroid Build Coastguard Worker __m128i vacc1 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (x + 8)));
36*4bdc9457SAndroid Build Coastguard Worker __m128i vacc2 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (x + 16)));
37*4bdc9457SAndroid Build Coastguard Worker __m128i vacc3 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (x + 24)));
38*4bdc9457SAndroid Build Coastguard Worker x += 32;
39*4bdc9457SAndroid Build Coastguard Worker
40*4bdc9457SAndroid Build Coastguard Worker vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
41*4bdc9457SAndroid Build Coastguard Worker vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
42*4bdc9457SAndroid Build Coastguard Worker vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
43*4bdc9457SAndroid Build Coastguard Worker vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
44*4bdc9457SAndroid Build Coastguard Worker
45*4bdc9457SAndroid Build Coastguard Worker vacc0 = _mm_slli_epi16(vacc0, 7);
46*4bdc9457SAndroid Build Coastguard Worker vacc1 = _mm_slli_epi16(vacc1, 7);
47*4bdc9457SAndroid Build Coastguard Worker vacc2 = _mm_slli_epi16(vacc2, 7);
48*4bdc9457SAndroid Build Coastguard Worker vacc3 = _mm_slli_epi16(vacc3, 7);
49*4bdc9457SAndroid Build Coastguard Worker
50*4bdc9457SAndroid Build Coastguard Worker vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
51*4bdc9457SAndroid Build Coastguard Worker vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
52*4bdc9457SAndroid Build Coastguard Worker vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier);
53*4bdc9457SAndroid Build Coastguard Worker vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier);
54*4bdc9457SAndroid Build Coastguard Worker
55*4bdc9457SAndroid Build Coastguard Worker vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
56*4bdc9457SAndroid Build Coastguard Worker vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
57*4bdc9457SAndroid Build Coastguard Worker vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
58*4bdc9457SAndroid Build Coastguard Worker vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
59*4bdc9457SAndroid Build Coastguard Worker
60*4bdc9457SAndroid Build Coastguard Worker const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
61*4bdc9457SAndroid Build Coastguard Worker const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
62*4bdc9457SAndroid Build Coastguard Worker
63*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si128((__m128i*) y, vy0);
64*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si128((__m128i*) (y + 16), vy1);
65*4bdc9457SAndroid Build Coastguard Worker y += 32;
66*4bdc9457SAndroid Build Coastguard Worker }
67*4bdc9457SAndroid Build Coastguard Worker for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
68*4bdc9457SAndroid Build Coastguard Worker __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
69*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_sub_epi16(vinput_zero_point, vacc);
70*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_slli_epi16(vacc, 7);
71*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
72*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_adds_epi16(vacc, voutput_zero_point);
73*4bdc9457SAndroid Build Coastguard Worker x += 8;
74*4bdc9457SAndroid Build Coastguard Worker
75*4bdc9457SAndroid Build Coastguard Worker const __m128i vy = _mm_packus_epi16(vacc, vacc);
76*4bdc9457SAndroid Build Coastguard Worker _mm_storel_epi64((__m128i*) y, vy);
77*4bdc9457SAndroid Build Coastguard Worker y += 8;
78*4bdc9457SAndroid Build Coastguard Worker }
79*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(n != 0) {
80*4bdc9457SAndroid Build Coastguard Worker assert(n >= 1 * sizeof(uint8_t));
81*4bdc9457SAndroid Build Coastguard Worker assert(n <= 7 * sizeof(uint8_t));
82*4bdc9457SAndroid Build Coastguard Worker
83*4bdc9457SAndroid Build Coastguard Worker __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
84*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_sub_epi16(vinput_zero_point, vacc);
85*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_slli_epi16(vacc, 7);
86*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
87*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_adds_epi16(vacc, voutput_zero_point);
88*4bdc9457SAndroid Build Coastguard Worker
89*4bdc9457SAndroid Build Coastguard Worker __m128i vy = _mm_packus_epi16(vacc, vacc);
90*4bdc9457SAndroid Build Coastguard Worker if (n & (4 * sizeof(uint8_t))) {
91*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si32(y, vy);
92*4bdc9457SAndroid Build Coastguard Worker vy = _mm_srli_epi64(vy, 32);
93*4bdc9457SAndroid Build Coastguard Worker y += 4;
94*4bdc9457SAndroid Build Coastguard Worker }
95*4bdc9457SAndroid Build Coastguard Worker if (n & (2 * sizeof(uint8_t))) {
96*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si16(y, vy);
97*4bdc9457SAndroid Build Coastguard Worker vy = _mm_srli_epi32(vy, 16);
98*4bdc9457SAndroid Build Coastguard Worker y += 2;
99*4bdc9457SAndroid Build Coastguard Worker }
100*4bdc9457SAndroid Build Coastguard Worker if (n & (1 * sizeof(uint8_t))) {
101*4bdc9457SAndroid Build Coastguard Worker *y = (uint8_t) _mm_extract_epi8(vy, 0);
102*4bdc9457SAndroid Build Coastguard Worker }
103*4bdc9457SAndroid Build Coastguard Worker }
104*4bdc9457SAndroid Build Coastguard Worker }
105