1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/qs8-dwconv/unipass-avx2-mul16-vpunpck.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 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/dwconv.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/unaligned.h>
16*4bdc9457SAndroid Build Coastguard Worker
17*4bdc9457SAndroid Build Coastguard Worker
xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18*4bdc9457SAndroid Build Coastguard Worker void xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck(
19*4bdc9457SAndroid Build Coastguard Worker size_t channels,
20*4bdc9457SAndroid Build Coastguard Worker size_t output_width,
21*4bdc9457SAndroid Build Coastguard Worker const int8_t** input,
22*4bdc9457SAndroid Build Coastguard Worker const void* weights,
23*4bdc9457SAndroid Build Coastguard Worker int8_t* output,
24*4bdc9457SAndroid Build Coastguard Worker size_t input_stride,
25*4bdc9457SAndroid Build Coastguard Worker size_t output_increment,
26*4bdc9457SAndroid Build Coastguard Worker size_t input_offset,
27*4bdc9457SAndroid Build Coastguard Worker const int8_t* zero,
28*4bdc9457SAndroid Build Coastguard Worker const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29*4bdc9457SAndroid Build Coastguard Worker {
30*4bdc9457SAndroid Build Coastguard Worker assert(channels != 0);
31*4bdc9457SAndroid Build Coastguard Worker assert(output_width != 0);
32*4bdc9457SAndroid Build Coastguard Worker
33*4bdc9457SAndroid Build Coastguard Worker do {
34*4bdc9457SAndroid Build Coastguard Worker const int8_t* i0 = input[0];
35*4bdc9457SAndroid Build Coastguard Worker assert(i0 != NULL);
36*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i0 != zero) {
37*4bdc9457SAndroid Build Coastguard Worker i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38*4bdc9457SAndroid Build Coastguard Worker }
39*4bdc9457SAndroid Build Coastguard Worker const int8_t* i1 = input[1];
40*4bdc9457SAndroid Build Coastguard Worker assert(i1 != NULL);
41*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i1 != zero) {
42*4bdc9457SAndroid Build Coastguard Worker i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43*4bdc9457SAndroid Build Coastguard Worker }
44*4bdc9457SAndroid Build Coastguard Worker const int8_t* i2 = input[2];
45*4bdc9457SAndroid Build Coastguard Worker assert(i2 != NULL);
46*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i2 != zero) {
47*4bdc9457SAndroid Build Coastguard Worker i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48*4bdc9457SAndroid Build Coastguard Worker }
49*4bdc9457SAndroid Build Coastguard Worker const int8_t* i3 = input[3];
50*4bdc9457SAndroid Build Coastguard Worker assert(i3 != NULL);
51*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i3 != zero) {
52*4bdc9457SAndroid Build Coastguard Worker i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53*4bdc9457SAndroid Build Coastguard Worker }
54*4bdc9457SAndroid Build Coastguard Worker const int8_t* i4 = input[4];
55*4bdc9457SAndroid Build Coastguard Worker assert(i4 != NULL);
56*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i4 != zero) {
57*4bdc9457SAndroid Build Coastguard Worker i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58*4bdc9457SAndroid Build Coastguard Worker }
59*4bdc9457SAndroid Build Coastguard Worker const int8_t* i5 = input[5];
60*4bdc9457SAndroid Build Coastguard Worker assert(i5 != NULL);
61*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i5 != zero) {
62*4bdc9457SAndroid Build Coastguard Worker i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63*4bdc9457SAndroid Build Coastguard Worker }
64*4bdc9457SAndroid Build Coastguard Worker const int8_t* i6 = input[6];
65*4bdc9457SAndroid Build Coastguard Worker assert(i6 != NULL);
66*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i6 != zero) {
67*4bdc9457SAndroid Build Coastguard Worker i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68*4bdc9457SAndroid Build Coastguard Worker }
69*4bdc9457SAndroid Build Coastguard Worker const int8_t* i7 = input[7];
70*4bdc9457SAndroid Build Coastguard Worker assert(i7 != NULL);
71*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i7 != zero) {
72*4bdc9457SAndroid Build Coastguard Worker i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73*4bdc9457SAndroid Build Coastguard Worker }
74*4bdc9457SAndroid Build Coastguard Worker const int8_t* i8 = input[8];
75*4bdc9457SAndroid Build Coastguard Worker assert(i8 != NULL);
76*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i8 != zero) {
77*4bdc9457SAndroid Build Coastguard Worker i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78*4bdc9457SAndroid Build Coastguard Worker }
79*4bdc9457SAndroid Build Coastguard Worker const int8_t* i9 = input[9];
80*4bdc9457SAndroid Build Coastguard Worker assert(i9 != NULL);
81*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i9 != zero) {
82*4bdc9457SAndroid Build Coastguard Worker i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
83*4bdc9457SAndroid Build Coastguard Worker }
84*4bdc9457SAndroid Build Coastguard Worker const int8_t* i10 = input[10];
85*4bdc9457SAndroid Build Coastguard Worker assert(i10 != NULL);
86*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i10 != zero) {
87*4bdc9457SAndroid Build Coastguard Worker i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
88*4bdc9457SAndroid Build Coastguard Worker }
89*4bdc9457SAndroid Build Coastguard Worker const int8_t* i11 = input[11];
90*4bdc9457SAndroid Build Coastguard Worker assert(i11 != NULL);
91*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i11 != zero) {
92*4bdc9457SAndroid Build Coastguard Worker i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
93*4bdc9457SAndroid Build Coastguard Worker }
94*4bdc9457SAndroid Build Coastguard Worker const int8_t* i12 = input[12];
95*4bdc9457SAndroid Build Coastguard Worker assert(i12 != NULL);
96*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i12 != zero) {
97*4bdc9457SAndroid Build Coastguard Worker i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
98*4bdc9457SAndroid Build Coastguard Worker }
99*4bdc9457SAndroid Build Coastguard Worker const int8_t* i13 = input[13];
100*4bdc9457SAndroid Build Coastguard Worker assert(i13 != NULL);
101*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i13 != zero) {
102*4bdc9457SAndroid Build Coastguard Worker i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
103*4bdc9457SAndroid Build Coastguard Worker }
104*4bdc9457SAndroid Build Coastguard Worker const int8_t* i14 = input[14];
105*4bdc9457SAndroid Build Coastguard Worker assert(i14 != NULL);
106*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i14 != zero) {
107*4bdc9457SAndroid Build Coastguard Worker i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
108*4bdc9457SAndroid Build Coastguard Worker }
109*4bdc9457SAndroid Build Coastguard Worker const int8_t* i15 = input[15];
110*4bdc9457SAndroid Build Coastguard Worker assert(i15 != NULL);
111*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i15 != zero) {
112*4bdc9457SAndroid Build Coastguard Worker i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
113*4bdc9457SAndroid Build Coastguard Worker }
114*4bdc9457SAndroid Build Coastguard Worker const int8_t* i16 = input[16];
115*4bdc9457SAndroid Build Coastguard Worker assert(i16 != NULL);
116*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i16 != zero) {
117*4bdc9457SAndroid Build Coastguard Worker i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
118*4bdc9457SAndroid Build Coastguard Worker }
119*4bdc9457SAndroid Build Coastguard Worker const int8_t* i17 = input[17];
120*4bdc9457SAndroid Build Coastguard Worker assert(i17 != NULL);
121*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i17 != zero) {
122*4bdc9457SAndroid Build Coastguard Worker i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
123*4bdc9457SAndroid Build Coastguard Worker }
124*4bdc9457SAndroid Build Coastguard Worker const int8_t* i18 = input[18];
125*4bdc9457SAndroid Build Coastguard Worker assert(i18 != NULL);
126*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i18 != zero) {
127*4bdc9457SAndroid Build Coastguard Worker i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
128*4bdc9457SAndroid Build Coastguard Worker }
129*4bdc9457SAndroid Build Coastguard Worker const int8_t* i19 = input[19];
130*4bdc9457SAndroid Build Coastguard Worker assert(i19 != NULL);
131*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i19 != zero) {
132*4bdc9457SAndroid Build Coastguard Worker i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
133*4bdc9457SAndroid Build Coastguard Worker }
134*4bdc9457SAndroid Build Coastguard Worker const int8_t* i20 = input[20];
135*4bdc9457SAndroid Build Coastguard Worker assert(i20 != NULL);
136*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i20 != zero) {
137*4bdc9457SAndroid Build Coastguard Worker i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
138*4bdc9457SAndroid Build Coastguard Worker }
139*4bdc9457SAndroid Build Coastguard Worker const int8_t* i21 = input[21];
140*4bdc9457SAndroid Build Coastguard Worker assert(i21 != NULL);
141*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i21 != zero) {
142*4bdc9457SAndroid Build Coastguard Worker i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
143*4bdc9457SAndroid Build Coastguard Worker }
144*4bdc9457SAndroid Build Coastguard Worker const int8_t* i22 = input[22];
145*4bdc9457SAndroid Build Coastguard Worker assert(i22 != NULL);
146*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i22 != zero) {
147*4bdc9457SAndroid Build Coastguard Worker i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
148*4bdc9457SAndroid Build Coastguard Worker }
149*4bdc9457SAndroid Build Coastguard Worker const int8_t* i23 = input[23];
150*4bdc9457SAndroid Build Coastguard Worker assert(i23 != NULL);
151*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i23 != zero) {
152*4bdc9457SAndroid Build Coastguard Worker i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
153*4bdc9457SAndroid Build Coastguard Worker }
154*4bdc9457SAndroid Build Coastguard Worker const int8_t* i24 = input[24];
155*4bdc9457SAndroid Build Coastguard Worker assert(i24 != NULL);
156*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i24 != zero) {
157*4bdc9457SAndroid Build Coastguard Worker i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
158*4bdc9457SAndroid Build Coastguard Worker }
159*4bdc9457SAndroid Build Coastguard Worker input = (const int8_t**) ((uintptr_t) input + input_stride);
160*4bdc9457SAndroid Build Coastguard Worker
161*4bdc9457SAndroid Build Coastguard Worker size_t c = channels;
162*4bdc9457SAndroid Build Coastguard Worker const void* w = weights;
163*4bdc9457SAndroid Build Coastguard Worker for (; c >= 32; c -= 32) {
164*4bdc9457SAndroid Build Coastguard Worker __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
165*4bdc9457SAndroid Build Coastguard Worker __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
166*4bdc9457SAndroid Build Coastguard Worker __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int32_t)));
167*4bdc9457SAndroid Build Coastguard Worker __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 24 * sizeof(int32_t)));
168*4bdc9457SAndroid Build Coastguard Worker
169*4bdc9457SAndroid Build Coastguard Worker __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
170*4bdc9457SAndroid Build Coastguard Worker __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
171*4bdc9457SAndroid Build Coastguard Worker __m256i vaccGHIJOPQR = _mm256_inserti128_si256(vaccGHIJKLMN, _mm256_castsi256_si128(vaccOPQRSTUV), 1);
172*4bdc9457SAndroid Build Coastguard Worker __m256i vaccKLMNSTUV = _mm256_permute2x128_si256(vaccGHIJKLMN, vaccOPQRSTUV, 0x31);
173*4bdc9457SAndroid Build Coastguard Worker
174*4bdc9457SAndroid Build Coastguard Worker
175*4bdc9457SAndroid Build Coastguard Worker const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
176*4bdc9457SAndroid Build Coastguard Worker const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
177*4bdc9457SAndroid Build Coastguard Worker const __m256i vi0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i0 + 16)));
178*4bdc9457SAndroid Build Coastguard Worker const __m256i vk0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
179*4bdc9457SAndroid Build Coastguard Worker i0 += 32;
180*4bdc9457SAndroid Build Coastguard Worker
181*4bdc9457SAndroid Build Coastguard Worker __m256i vacc0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
182*4bdc9457SAndroid Build Coastguard Worker __m256i vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV);
183*4bdc9457SAndroid Build Coastguard Worker
184*4bdc9457SAndroid Build Coastguard Worker
185*4bdc9457SAndroid Build Coastguard Worker const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
186*4bdc9457SAndroid Build Coastguard Worker const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
187*4bdc9457SAndroid Build Coastguard Worker const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16)));
188*4bdc9457SAndroid Build Coastguard Worker const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
189*4bdc9457SAndroid Build Coastguard Worker i1 += 32;
190*4bdc9457SAndroid Build Coastguard Worker
191*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
192*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV));
193*4bdc9457SAndroid Build Coastguard Worker
194*4bdc9457SAndroid Build Coastguard Worker __m256i vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
195*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
196*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
197*4bdc9457SAndroid Build Coastguard Worker __m256i vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
198*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
199*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
200*4bdc9457SAndroid Build Coastguard Worker
201*4bdc9457SAndroid Build Coastguard Worker const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
202*4bdc9457SAndroid Build Coastguard Worker const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
203*4bdc9457SAndroid Build Coastguard Worker const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16)));
204*4bdc9457SAndroid Build Coastguard Worker const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
205*4bdc9457SAndroid Build Coastguard Worker i2 += 32;
206*4bdc9457SAndroid Build Coastguard Worker
207*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
208*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV);
209*4bdc9457SAndroid Build Coastguard Worker
210*4bdc9457SAndroid Build Coastguard Worker
211*4bdc9457SAndroid Build Coastguard Worker const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
212*4bdc9457SAndroid Build Coastguard Worker const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
213*4bdc9457SAndroid Build Coastguard Worker const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16)));
214*4bdc9457SAndroid Build Coastguard Worker const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
215*4bdc9457SAndroid Build Coastguard Worker i3 += 32;
216*4bdc9457SAndroid Build Coastguard Worker
217*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
218*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV));
219*4bdc9457SAndroid Build Coastguard Worker
220*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
221*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
222*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
223*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
224*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
225*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
226*4bdc9457SAndroid Build Coastguard Worker
227*4bdc9457SAndroid Build Coastguard Worker const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
228*4bdc9457SAndroid Build Coastguard Worker const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
229*4bdc9457SAndroid Build Coastguard Worker const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16)));
230*4bdc9457SAndroid Build Coastguard Worker const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
231*4bdc9457SAndroid Build Coastguard Worker i4 += 32;
232*4bdc9457SAndroid Build Coastguard Worker
233*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
234*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV);
235*4bdc9457SAndroid Build Coastguard Worker
236*4bdc9457SAndroid Build Coastguard Worker
237*4bdc9457SAndroid Build Coastguard Worker const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
238*4bdc9457SAndroid Build Coastguard Worker const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
239*4bdc9457SAndroid Build Coastguard Worker const __m256i vi5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i5 + 16)));
240*4bdc9457SAndroid Build Coastguard Worker const __m256i vk5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
241*4bdc9457SAndroid Build Coastguard Worker i5 += 32;
242*4bdc9457SAndroid Build Coastguard Worker
243*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
244*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV));
245*4bdc9457SAndroid Build Coastguard Worker
246*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
247*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
248*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
249*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
250*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
251*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
252*4bdc9457SAndroid Build Coastguard Worker
253*4bdc9457SAndroid Build Coastguard Worker const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
254*4bdc9457SAndroid Build Coastguard Worker const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
255*4bdc9457SAndroid Build Coastguard Worker const __m256i vi6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i6 + 16)));
256*4bdc9457SAndroid Build Coastguard Worker const __m256i vk6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
257*4bdc9457SAndroid Build Coastguard Worker i6 += 32;
258*4bdc9457SAndroid Build Coastguard Worker
259*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
260*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV);
261*4bdc9457SAndroid Build Coastguard Worker
262*4bdc9457SAndroid Build Coastguard Worker
263*4bdc9457SAndroid Build Coastguard Worker const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
264*4bdc9457SAndroid Build Coastguard Worker const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
265*4bdc9457SAndroid Build Coastguard Worker const __m256i vi7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i7 + 16)));
266*4bdc9457SAndroid Build Coastguard Worker const __m256i vk7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
267*4bdc9457SAndroid Build Coastguard Worker i7 += 32;
268*4bdc9457SAndroid Build Coastguard Worker
269*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
270*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV));
271*4bdc9457SAndroid Build Coastguard Worker
272*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
273*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
274*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
275*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
276*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
277*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
278*4bdc9457SAndroid Build Coastguard Worker
279*4bdc9457SAndroid Build Coastguard Worker const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
280*4bdc9457SAndroid Build Coastguard Worker const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
281*4bdc9457SAndroid Build Coastguard Worker const __m256i vi8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i8 + 16)));
282*4bdc9457SAndroid Build Coastguard Worker const __m256i vk8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
283*4bdc9457SAndroid Build Coastguard Worker i8 += 32;
284*4bdc9457SAndroid Build Coastguard Worker
285*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
286*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV);
287*4bdc9457SAndroid Build Coastguard Worker
288*4bdc9457SAndroid Build Coastguard Worker
289*4bdc9457SAndroid Build Coastguard Worker const __m256i vi9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i9));
290*4bdc9457SAndroid Build Coastguard Worker const __m256i vk9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t))));
291*4bdc9457SAndroid Build Coastguard Worker const __m256i vi9xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i9 + 16)));
292*4bdc9457SAndroid Build Coastguard Worker const __m256i vk9xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 304 * sizeof(int8_t))));
293*4bdc9457SAndroid Build Coastguard Worker i9 += 32;
294*4bdc9457SAndroid Build Coastguard Worker
295*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF));
296*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi9xGHIJKLMNOPQRSTUV, vk9xGHIJKLMNOPQRSTUV));
297*4bdc9457SAndroid Build Coastguard Worker
298*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
299*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
300*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
301*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
302*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
303*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
304*4bdc9457SAndroid Build Coastguard Worker
305*4bdc9457SAndroid Build Coastguard Worker const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10));
306*4bdc9457SAndroid Build Coastguard Worker const __m256i vk10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 320 * sizeof(int8_t))));
307*4bdc9457SAndroid Build Coastguard Worker const __m256i vi10xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i10 + 16)));
308*4bdc9457SAndroid Build Coastguard Worker const __m256i vk10xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 336 * sizeof(int8_t))));
309*4bdc9457SAndroid Build Coastguard Worker i10 += 32;
310*4bdc9457SAndroid Build Coastguard Worker
311*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF);
312*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi10xGHIJKLMNOPQRSTUV, vk10xGHIJKLMNOPQRSTUV);
313*4bdc9457SAndroid Build Coastguard Worker
314*4bdc9457SAndroid Build Coastguard Worker
315*4bdc9457SAndroid Build Coastguard Worker const __m256i vi11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i11));
316*4bdc9457SAndroid Build Coastguard Worker const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 352 * sizeof(int8_t))));
317*4bdc9457SAndroid Build Coastguard Worker const __m256i vi11xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i11 + 16)));
318*4bdc9457SAndroid Build Coastguard Worker const __m256i vk11xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 368 * sizeof(int8_t))));
319*4bdc9457SAndroid Build Coastguard Worker i11 += 32;
320*4bdc9457SAndroid Build Coastguard Worker
321*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF));
322*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi11xGHIJKLMNOPQRSTUV, vk11xGHIJKLMNOPQRSTUV));
323*4bdc9457SAndroid Build Coastguard Worker
324*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
325*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
326*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
327*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
328*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
329*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
330*4bdc9457SAndroid Build Coastguard Worker
331*4bdc9457SAndroid Build Coastguard Worker const __m256i vi12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i12));
332*4bdc9457SAndroid Build Coastguard Worker const __m256i vk12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 384 * sizeof(int8_t))));
333*4bdc9457SAndroid Build Coastguard Worker const __m256i vi12xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i12 + 16)));
334*4bdc9457SAndroid Build Coastguard Worker const __m256i vk12xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 400 * sizeof(int8_t))));
335*4bdc9457SAndroid Build Coastguard Worker i12 += 32;
336*4bdc9457SAndroid Build Coastguard Worker
337*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF);
338*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi12xGHIJKLMNOPQRSTUV, vk12xGHIJKLMNOPQRSTUV);
339*4bdc9457SAndroid Build Coastguard Worker
340*4bdc9457SAndroid Build Coastguard Worker
341*4bdc9457SAndroid Build Coastguard Worker const __m256i vi13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i13));
342*4bdc9457SAndroid Build Coastguard Worker const __m256i vk13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 416 * sizeof(int8_t))));
343*4bdc9457SAndroid Build Coastguard Worker const __m256i vi13xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i13 + 16)));
344*4bdc9457SAndroid Build Coastguard Worker const __m256i vk13xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 432 * sizeof(int8_t))));
345*4bdc9457SAndroid Build Coastguard Worker i13 += 32;
346*4bdc9457SAndroid Build Coastguard Worker
347*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF));
348*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi13xGHIJKLMNOPQRSTUV, vk13xGHIJKLMNOPQRSTUV));
349*4bdc9457SAndroid Build Coastguard Worker
350*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
351*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
352*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
353*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
354*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
355*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
356*4bdc9457SAndroid Build Coastguard Worker
357*4bdc9457SAndroid Build Coastguard Worker const __m256i vi14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i14));
358*4bdc9457SAndroid Build Coastguard Worker const __m256i vk14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 448 * sizeof(int8_t))));
359*4bdc9457SAndroid Build Coastguard Worker const __m256i vi14xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i14 + 16)));
360*4bdc9457SAndroid Build Coastguard Worker const __m256i vk14xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 464 * sizeof(int8_t))));
361*4bdc9457SAndroid Build Coastguard Worker i14 += 32;
362*4bdc9457SAndroid Build Coastguard Worker
363*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF);
364*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi14xGHIJKLMNOPQRSTUV, vk14xGHIJKLMNOPQRSTUV);
365*4bdc9457SAndroid Build Coastguard Worker
366*4bdc9457SAndroid Build Coastguard Worker
367*4bdc9457SAndroid Build Coastguard Worker const __m256i vi15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i15));
368*4bdc9457SAndroid Build Coastguard Worker const __m256i vk15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 480 * sizeof(int8_t))));
369*4bdc9457SAndroid Build Coastguard Worker const __m256i vi15xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i15 + 16)));
370*4bdc9457SAndroid Build Coastguard Worker const __m256i vk15xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 496 * sizeof(int8_t))));
371*4bdc9457SAndroid Build Coastguard Worker i15 += 32;
372*4bdc9457SAndroid Build Coastguard Worker
373*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF));
374*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi15xGHIJKLMNOPQRSTUV, vk15xGHIJKLMNOPQRSTUV));
375*4bdc9457SAndroid Build Coastguard Worker
376*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
377*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
378*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
379*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
380*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
381*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
382*4bdc9457SAndroid Build Coastguard Worker
383*4bdc9457SAndroid Build Coastguard Worker const __m256i vi16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i16));
384*4bdc9457SAndroid Build Coastguard Worker const __m256i vk16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 512 * sizeof(int8_t))));
385*4bdc9457SAndroid Build Coastguard Worker const __m256i vi16xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i16 + 16)));
386*4bdc9457SAndroid Build Coastguard Worker const __m256i vk16xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 528 * sizeof(int8_t))));
387*4bdc9457SAndroid Build Coastguard Worker i16 += 32;
388*4bdc9457SAndroid Build Coastguard Worker
389*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF);
390*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi16xGHIJKLMNOPQRSTUV, vk16xGHIJKLMNOPQRSTUV);
391*4bdc9457SAndroid Build Coastguard Worker
392*4bdc9457SAndroid Build Coastguard Worker
393*4bdc9457SAndroid Build Coastguard Worker const __m256i vi17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i17));
394*4bdc9457SAndroid Build Coastguard Worker const __m256i vk17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 544 * sizeof(int8_t))));
395*4bdc9457SAndroid Build Coastguard Worker const __m256i vi17xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i17 + 16)));
396*4bdc9457SAndroid Build Coastguard Worker const __m256i vk17xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 560 * sizeof(int8_t))));
397*4bdc9457SAndroid Build Coastguard Worker i17 += 32;
398*4bdc9457SAndroid Build Coastguard Worker
399*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF));
400*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi17xGHIJKLMNOPQRSTUV, vk17xGHIJKLMNOPQRSTUV));
401*4bdc9457SAndroid Build Coastguard Worker
402*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
403*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
404*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
405*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
406*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
407*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
408*4bdc9457SAndroid Build Coastguard Worker
409*4bdc9457SAndroid Build Coastguard Worker const __m256i vi18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i18));
410*4bdc9457SAndroid Build Coastguard Worker const __m256i vk18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 576 * sizeof(int8_t))));
411*4bdc9457SAndroid Build Coastguard Worker const __m256i vi18xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i18 + 16)));
412*4bdc9457SAndroid Build Coastguard Worker const __m256i vk18xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 592 * sizeof(int8_t))));
413*4bdc9457SAndroid Build Coastguard Worker i18 += 32;
414*4bdc9457SAndroid Build Coastguard Worker
415*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF);
416*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi18xGHIJKLMNOPQRSTUV, vk18xGHIJKLMNOPQRSTUV);
417*4bdc9457SAndroid Build Coastguard Worker
418*4bdc9457SAndroid Build Coastguard Worker
419*4bdc9457SAndroid Build Coastguard Worker const __m256i vi19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i19));
420*4bdc9457SAndroid Build Coastguard Worker const __m256i vk19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 608 * sizeof(int8_t))));
421*4bdc9457SAndroid Build Coastguard Worker const __m256i vi19xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i19 + 16)));
422*4bdc9457SAndroid Build Coastguard Worker const __m256i vk19xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 624 * sizeof(int8_t))));
423*4bdc9457SAndroid Build Coastguard Worker i19 += 32;
424*4bdc9457SAndroid Build Coastguard Worker
425*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF));
426*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi19xGHIJKLMNOPQRSTUV, vk19xGHIJKLMNOPQRSTUV));
427*4bdc9457SAndroid Build Coastguard Worker
428*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
429*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
430*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
431*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
432*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
433*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
434*4bdc9457SAndroid Build Coastguard Worker
435*4bdc9457SAndroid Build Coastguard Worker const __m256i vi20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i20));
436*4bdc9457SAndroid Build Coastguard Worker const __m256i vk20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 640 * sizeof(int8_t))));
437*4bdc9457SAndroid Build Coastguard Worker const __m256i vi20xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i20 + 16)));
438*4bdc9457SAndroid Build Coastguard Worker const __m256i vk20xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 656 * sizeof(int8_t))));
439*4bdc9457SAndroid Build Coastguard Worker i20 += 32;
440*4bdc9457SAndroid Build Coastguard Worker
441*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF);
442*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi20xGHIJKLMNOPQRSTUV, vk20xGHIJKLMNOPQRSTUV);
443*4bdc9457SAndroid Build Coastguard Worker
444*4bdc9457SAndroid Build Coastguard Worker
445*4bdc9457SAndroid Build Coastguard Worker const __m256i vi21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i21));
446*4bdc9457SAndroid Build Coastguard Worker const __m256i vk21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 672 * sizeof(int8_t))));
447*4bdc9457SAndroid Build Coastguard Worker const __m256i vi21xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i21 + 16)));
448*4bdc9457SAndroid Build Coastguard Worker const __m256i vk21xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 688 * sizeof(int8_t))));
449*4bdc9457SAndroid Build Coastguard Worker i21 += 32;
450*4bdc9457SAndroid Build Coastguard Worker
451*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF));
452*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi21xGHIJKLMNOPQRSTUV, vk21xGHIJKLMNOPQRSTUV));
453*4bdc9457SAndroid Build Coastguard Worker
454*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
455*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
456*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
457*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
458*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
459*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
460*4bdc9457SAndroid Build Coastguard Worker
461*4bdc9457SAndroid Build Coastguard Worker const __m256i vi22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i22));
462*4bdc9457SAndroid Build Coastguard Worker const __m256i vk22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 704 * sizeof(int8_t))));
463*4bdc9457SAndroid Build Coastguard Worker const __m256i vi22xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i22 + 16)));
464*4bdc9457SAndroid Build Coastguard Worker const __m256i vk22xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 720 * sizeof(int8_t))));
465*4bdc9457SAndroid Build Coastguard Worker i22 += 32;
466*4bdc9457SAndroid Build Coastguard Worker
467*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF);
468*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi22xGHIJKLMNOPQRSTUV, vk22xGHIJKLMNOPQRSTUV);
469*4bdc9457SAndroid Build Coastguard Worker
470*4bdc9457SAndroid Build Coastguard Worker
471*4bdc9457SAndroid Build Coastguard Worker const __m256i vi23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i23));
472*4bdc9457SAndroid Build Coastguard Worker const __m256i vk23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 736 * sizeof(int8_t))));
473*4bdc9457SAndroid Build Coastguard Worker const __m256i vi23xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i23 + 16)));
474*4bdc9457SAndroid Build Coastguard Worker const __m256i vk23xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 752 * sizeof(int8_t))));
475*4bdc9457SAndroid Build Coastguard Worker i23 += 32;
476*4bdc9457SAndroid Build Coastguard Worker
477*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF));
478*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi23xGHIJKLMNOPQRSTUV, vk23xGHIJKLMNOPQRSTUV));
479*4bdc9457SAndroid Build Coastguard Worker
480*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
481*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
482*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
483*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
484*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
485*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
486*4bdc9457SAndroid Build Coastguard Worker
487*4bdc9457SAndroid Build Coastguard Worker const __m256i vi24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i24));
488*4bdc9457SAndroid Build Coastguard Worker const __m256i vk24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 768 * sizeof(int8_t))));
489*4bdc9457SAndroid Build Coastguard Worker const __m256i vi24xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i24 + 16)));
490*4bdc9457SAndroid Build Coastguard Worker const __m256i vk24xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 784 * sizeof(int8_t))));
491*4bdc9457SAndroid Build Coastguard Worker i24 += 32;
492*4bdc9457SAndroid Build Coastguard Worker
493*4bdc9457SAndroid Build Coastguard Worker vacc0123456789ABCDEF = _mm256_mullo_epi16(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF);
494*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi24xGHIJKLMNOPQRSTUV, vk24xGHIJKLMNOPQRSTUV);
495*4bdc9457SAndroid Build Coastguard Worker
496*4bdc9457SAndroid Build Coastguard Worker vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
497*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
498*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
499*4bdc9457SAndroid Build Coastguard Worker vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
500*4bdc9457SAndroid Build Coastguard Worker vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
501*4bdc9457SAndroid Build Coastguard Worker vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
502*4bdc9457SAndroid Build Coastguard Worker
503*4bdc9457SAndroid Build Coastguard Worker w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t));
504*4bdc9457SAndroid Build Coastguard Worker
505*4bdc9457SAndroid Build Coastguard Worker vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
506*4bdc9457SAndroid Build Coastguard Worker vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
507*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMN = _mm256_inserti128_si256(vaccGHIJOPQR, _mm256_castsi256_si128(vaccKLMNSTUV), 1);
508*4bdc9457SAndroid Build Coastguard Worker vaccOPQRSTUV = _mm256_permute2x128_si256(vaccGHIJOPQR, vaccKLMNSTUV, 0x31);
509*4bdc9457SAndroid Build Coastguard Worker
510*4bdc9457SAndroid Build Coastguard Worker __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
511*4bdc9457SAndroid Build Coastguard Worker __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
512*4bdc9457SAndroid Build Coastguard Worker __m256 vfpaccGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
513*4bdc9457SAndroid Build Coastguard Worker __m256 vfpaccOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
514*4bdc9457SAndroid Build Coastguard Worker
515*4bdc9457SAndroid Build Coastguard Worker const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w);
516*4bdc9457SAndroid Build Coastguard Worker const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float)));
517*4bdc9457SAndroid Build Coastguard Worker const __m256 vscaleGHIJKLMN = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(float)));
518*4bdc9457SAndroid Build Coastguard Worker const __m256 vscaleOPQRSTUV = _mm256_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(float)));
519*4bdc9457SAndroid Build Coastguard Worker w = (const void*) ((uintptr_t) w + 32 * sizeof(float));
520*4bdc9457SAndroid Build Coastguard Worker vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
521*4bdc9457SAndroid Build Coastguard Worker vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
522*4bdc9457SAndroid Build Coastguard Worker vfpaccGHIJKLMN = _mm256_mul_ps(vfpaccGHIJKLMN, vscaleGHIJKLMN);
523*4bdc9457SAndroid Build Coastguard Worker vfpaccOPQRSTUV = _mm256_mul_ps(vfpaccOPQRSTUV, vscaleOPQRSTUV);
524*4bdc9457SAndroid Build Coastguard Worker
525*4bdc9457SAndroid Build Coastguard Worker const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
526*4bdc9457SAndroid Build Coastguard Worker vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
527*4bdc9457SAndroid Build Coastguard Worker vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
528*4bdc9457SAndroid Build Coastguard Worker vfpaccGHIJKLMN = _mm256_min_ps(vfpaccGHIJKLMN, voutput_max_less_zero_point);
529*4bdc9457SAndroid Build Coastguard Worker vfpaccOPQRSTUV = _mm256_min_ps(vfpaccOPQRSTUV, voutput_max_less_zero_point);
530*4bdc9457SAndroid Build Coastguard Worker
531*4bdc9457SAndroid Build Coastguard Worker vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
532*4bdc9457SAndroid Build Coastguard Worker vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
533*4bdc9457SAndroid Build Coastguard Worker vaccGHIJKLMN = _mm256_cvtps_epi32(vfpaccGHIJKLMN);
534*4bdc9457SAndroid Build Coastguard Worker vaccOPQRSTUV = _mm256_cvtps_epi32(vfpaccOPQRSTUV);
535*4bdc9457SAndroid Build Coastguard Worker
536*4bdc9457SAndroid Build Coastguard Worker const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
537*4bdc9457SAndroid Build Coastguard Worker const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
538*4bdc9457SAndroid Build Coastguard Worker const __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
539*4bdc9457SAndroid Build Coastguard Worker
540*4bdc9457SAndroid Build Coastguard Worker __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
541*4bdc9457SAndroid Build Coastguard Worker __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
542*4bdc9457SAndroid Build Coastguard Worker
543*4bdc9457SAndroid Build Coastguard Worker const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
544*4bdc9457SAndroid Build Coastguard Worker vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
545*4bdc9457SAndroid Build Coastguard Worker voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
546*4bdc9457SAndroid Build Coastguard Worker
547*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
548*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
549*4bdc9457SAndroid Build Coastguard Worker output += 32;
550*4bdc9457SAndroid Build Coastguard Worker }
551*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(c != 0) {
552*4bdc9457SAndroid Build Coastguard Worker const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t));
553*4bdc9457SAndroid Build Coastguard Worker do {
554*4bdc9457SAndroid Build Coastguard Worker __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
555*4bdc9457SAndroid Build Coastguard Worker __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
556*4bdc9457SAndroid Build Coastguard Worker
557*4bdc9457SAndroid Build Coastguard Worker __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
558*4bdc9457SAndroid Build Coastguard Worker __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
559*4bdc9457SAndroid Build Coastguard Worker
560*4bdc9457SAndroid Build Coastguard Worker
561*4bdc9457SAndroid Build Coastguard Worker const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
562*4bdc9457SAndroid Build Coastguard Worker const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k));
563*4bdc9457SAndroid Build Coastguard Worker i0 += 16;
564*4bdc9457SAndroid Build Coastguard Worker
565*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
566*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
567*4bdc9457SAndroid Build Coastguard Worker
568*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
569*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
570*4bdc9457SAndroid Build Coastguard Worker
571*4bdc9457SAndroid Build Coastguard Worker const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
572*4bdc9457SAndroid Build Coastguard Worker const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 32)));
573*4bdc9457SAndroid Build Coastguard Worker i1 += 16;
574*4bdc9457SAndroid Build Coastguard Worker
575*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
576*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
577*4bdc9457SAndroid Build Coastguard Worker
578*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
579*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
580*4bdc9457SAndroid Build Coastguard Worker
581*4bdc9457SAndroid Build Coastguard Worker const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
582*4bdc9457SAndroid Build Coastguard Worker const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 64)));
583*4bdc9457SAndroid Build Coastguard Worker i2 += 16;
584*4bdc9457SAndroid Build Coastguard Worker
585*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
586*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
587*4bdc9457SAndroid Build Coastguard Worker
588*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
589*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
590*4bdc9457SAndroid Build Coastguard Worker
591*4bdc9457SAndroid Build Coastguard Worker const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
592*4bdc9457SAndroid Build Coastguard Worker const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 96)));
593*4bdc9457SAndroid Build Coastguard Worker i3 += 16;
594*4bdc9457SAndroid Build Coastguard Worker
595*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
596*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
597*4bdc9457SAndroid Build Coastguard Worker
598*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
599*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
600*4bdc9457SAndroid Build Coastguard Worker
601*4bdc9457SAndroid Build Coastguard Worker const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
602*4bdc9457SAndroid Build Coastguard Worker const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 128)));
603*4bdc9457SAndroid Build Coastguard Worker i4 += 16;
604*4bdc9457SAndroid Build Coastguard Worker
605*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
606*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
607*4bdc9457SAndroid Build Coastguard Worker
608*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
609*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
610*4bdc9457SAndroid Build Coastguard Worker
611*4bdc9457SAndroid Build Coastguard Worker const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
612*4bdc9457SAndroid Build Coastguard Worker const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 160)));
613*4bdc9457SAndroid Build Coastguard Worker i5 += 16;
614*4bdc9457SAndroid Build Coastguard Worker
615*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
616*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
617*4bdc9457SAndroid Build Coastguard Worker
618*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
619*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
620*4bdc9457SAndroid Build Coastguard Worker
621*4bdc9457SAndroid Build Coastguard Worker const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
622*4bdc9457SAndroid Build Coastguard Worker const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 192)));
623*4bdc9457SAndroid Build Coastguard Worker i6 += 16;
624*4bdc9457SAndroid Build Coastguard Worker
625*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
626*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
627*4bdc9457SAndroid Build Coastguard Worker
628*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
629*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
630*4bdc9457SAndroid Build Coastguard Worker
631*4bdc9457SAndroid Build Coastguard Worker const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
632*4bdc9457SAndroid Build Coastguard Worker const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 224)));
633*4bdc9457SAndroid Build Coastguard Worker i7 += 16;
634*4bdc9457SAndroid Build Coastguard Worker
635*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod7x0123456789ABCDEFlo = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
636*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
637*4bdc9457SAndroid Build Coastguard Worker
638*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
639*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
640*4bdc9457SAndroid Build Coastguard Worker
641*4bdc9457SAndroid Build Coastguard Worker const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
642*4bdc9457SAndroid Build Coastguard Worker const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 256)));
643*4bdc9457SAndroid Build Coastguard Worker i8 += 16;
644*4bdc9457SAndroid Build Coastguard Worker
645*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod8x0123456789ABCDEFlo = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
646*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
647*4bdc9457SAndroid Build Coastguard Worker
648*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
649*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
650*4bdc9457SAndroid Build Coastguard Worker
651*4bdc9457SAndroid Build Coastguard Worker const __m256i vi9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i9));
652*4bdc9457SAndroid Build Coastguard Worker const __m256i vk9x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 288)));
653*4bdc9457SAndroid Build Coastguard Worker i9 += 16;
654*4bdc9457SAndroid Build Coastguard Worker
655*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod9x0123456789ABCDEFlo = _mm256_mullo_epi16(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF);
656*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod9x0123456789ABCDEFhi = _mm256_srai_epi16(vprod9x0123456789ABCDEFlo, 15);
657*4bdc9457SAndroid Build Coastguard Worker
658*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod9x0123456789ABCDEFlo, vprod9x0123456789ABCDEFhi));
659*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod9x0123456789ABCDEFlo, vprod9x0123456789ABCDEFhi));
660*4bdc9457SAndroid Build Coastguard Worker
661*4bdc9457SAndroid Build Coastguard Worker const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10));
662*4bdc9457SAndroid Build Coastguard Worker const __m256i vk10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 320)));
663*4bdc9457SAndroid Build Coastguard Worker i10 += 16;
664*4bdc9457SAndroid Build Coastguard Worker
665*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod10x0123456789ABCDEFlo = _mm256_mullo_epi16(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF);
666*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod10x0123456789ABCDEFhi = _mm256_srai_epi16(vprod10x0123456789ABCDEFlo, 15);
667*4bdc9457SAndroid Build Coastguard Worker
668*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod10x0123456789ABCDEFlo, vprod10x0123456789ABCDEFhi));
669*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod10x0123456789ABCDEFlo, vprod10x0123456789ABCDEFhi));
670*4bdc9457SAndroid Build Coastguard Worker
671*4bdc9457SAndroid Build Coastguard Worker const __m256i vi11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i11));
672*4bdc9457SAndroid Build Coastguard Worker const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 352)));
673*4bdc9457SAndroid Build Coastguard Worker i11 += 16;
674*4bdc9457SAndroid Build Coastguard Worker
675*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod11x0123456789ABCDEFlo = _mm256_mullo_epi16(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF);
676*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod11x0123456789ABCDEFhi = _mm256_srai_epi16(vprod11x0123456789ABCDEFlo, 15);
677*4bdc9457SAndroid Build Coastguard Worker
678*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod11x0123456789ABCDEFlo, vprod11x0123456789ABCDEFhi));
679*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod11x0123456789ABCDEFlo, vprod11x0123456789ABCDEFhi));
680*4bdc9457SAndroid Build Coastguard Worker
681*4bdc9457SAndroid Build Coastguard Worker const __m256i vi12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i12));
682*4bdc9457SAndroid Build Coastguard Worker const __m256i vk12x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 384)));
683*4bdc9457SAndroid Build Coastguard Worker i12 += 16;
684*4bdc9457SAndroid Build Coastguard Worker
685*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod12x0123456789ABCDEFlo = _mm256_mullo_epi16(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF);
686*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod12x0123456789ABCDEFhi = _mm256_srai_epi16(vprod12x0123456789ABCDEFlo, 15);
687*4bdc9457SAndroid Build Coastguard Worker
688*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod12x0123456789ABCDEFlo, vprod12x0123456789ABCDEFhi));
689*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod12x0123456789ABCDEFlo, vprod12x0123456789ABCDEFhi));
690*4bdc9457SAndroid Build Coastguard Worker
691*4bdc9457SAndroid Build Coastguard Worker const __m256i vi13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i13));
692*4bdc9457SAndroid Build Coastguard Worker const __m256i vk13x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 416)));
693*4bdc9457SAndroid Build Coastguard Worker i13 += 16;
694*4bdc9457SAndroid Build Coastguard Worker
695*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod13x0123456789ABCDEFlo = _mm256_mullo_epi16(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF);
696*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod13x0123456789ABCDEFhi = _mm256_srai_epi16(vprod13x0123456789ABCDEFlo, 15);
697*4bdc9457SAndroid Build Coastguard Worker
698*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod13x0123456789ABCDEFlo, vprod13x0123456789ABCDEFhi));
699*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod13x0123456789ABCDEFlo, vprod13x0123456789ABCDEFhi));
700*4bdc9457SAndroid Build Coastguard Worker
701*4bdc9457SAndroid Build Coastguard Worker const __m256i vi14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i14));
702*4bdc9457SAndroid Build Coastguard Worker const __m256i vk14x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 448)));
703*4bdc9457SAndroid Build Coastguard Worker i14 += 16;
704*4bdc9457SAndroid Build Coastguard Worker
705*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod14x0123456789ABCDEFlo = _mm256_mullo_epi16(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF);
706*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod14x0123456789ABCDEFhi = _mm256_srai_epi16(vprod14x0123456789ABCDEFlo, 15);
707*4bdc9457SAndroid Build Coastguard Worker
708*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod14x0123456789ABCDEFlo, vprod14x0123456789ABCDEFhi));
709*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod14x0123456789ABCDEFlo, vprod14x0123456789ABCDEFhi));
710*4bdc9457SAndroid Build Coastguard Worker
711*4bdc9457SAndroid Build Coastguard Worker const __m256i vi15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i15));
712*4bdc9457SAndroid Build Coastguard Worker const __m256i vk15x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 480)));
713*4bdc9457SAndroid Build Coastguard Worker i15 += 16;
714*4bdc9457SAndroid Build Coastguard Worker
715*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod15x0123456789ABCDEFlo = _mm256_mullo_epi16(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF);
716*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod15x0123456789ABCDEFhi = _mm256_srai_epi16(vprod15x0123456789ABCDEFlo, 15);
717*4bdc9457SAndroid Build Coastguard Worker
718*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod15x0123456789ABCDEFlo, vprod15x0123456789ABCDEFhi));
719*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod15x0123456789ABCDEFlo, vprod15x0123456789ABCDEFhi));
720*4bdc9457SAndroid Build Coastguard Worker
721*4bdc9457SAndroid Build Coastguard Worker const __m256i vi16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i16));
722*4bdc9457SAndroid Build Coastguard Worker const __m256i vk16x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 512)));
723*4bdc9457SAndroid Build Coastguard Worker i16 += 16;
724*4bdc9457SAndroid Build Coastguard Worker
725*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod16x0123456789ABCDEFlo = _mm256_mullo_epi16(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF);
726*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod16x0123456789ABCDEFhi = _mm256_srai_epi16(vprod16x0123456789ABCDEFlo, 15);
727*4bdc9457SAndroid Build Coastguard Worker
728*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod16x0123456789ABCDEFlo, vprod16x0123456789ABCDEFhi));
729*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod16x0123456789ABCDEFlo, vprod16x0123456789ABCDEFhi));
730*4bdc9457SAndroid Build Coastguard Worker
731*4bdc9457SAndroid Build Coastguard Worker const __m256i vi17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i17));
732*4bdc9457SAndroid Build Coastguard Worker const __m256i vk17x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 544)));
733*4bdc9457SAndroid Build Coastguard Worker i17 += 16;
734*4bdc9457SAndroid Build Coastguard Worker
735*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod17x0123456789ABCDEFlo = _mm256_mullo_epi16(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF);
736*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod17x0123456789ABCDEFhi = _mm256_srai_epi16(vprod17x0123456789ABCDEFlo, 15);
737*4bdc9457SAndroid Build Coastguard Worker
738*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod17x0123456789ABCDEFlo, vprod17x0123456789ABCDEFhi));
739*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod17x0123456789ABCDEFlo, vprod17x0123456789ABCDEFhi));
740*4bdc9457SAndroid Build Coastguard Worker
741*4bdc9457SAndroid Build Coastguard Worker const __m256i vi18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i18));
742*4bdc9457SAndroid Build Coastguard Worker const __m256i vk18x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 576)));
743*4bdc9457SAndroid Build Coastguard Worker i18 += 16;
744*4bdc9457SAndroid Build Coastguard Worker
745*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod18x0123456789ABCDEFlo = _mm256_mullo_epi16(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF);
746*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod18x0123456789ABCDEFhi = _mm256_srai_epi16(vprod18x0123456789ABCDEFlo, 15);
747*4bdc9457SAndroid Build Coastguard Worker
748*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod18x0123456789ABCDEFlo, vprod18x0123456789ABCDEFhi));
749*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod18x0123456789ABCDEFlo, vprod18x0123456789ABCDEFhi));
750*4bdc9457SAndroid Build Coastguard Worker
751*4bdc9457SAndroid Build Coastguard Worker const __m256i vi19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i19));
752*4bdc9457SAndroid Build Coastguard Worker const __m256i vk19x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 608)));
753*4bdc9457SAndroid Build Coastguard Worker i19 += 16;
754*4bdc9457SAndroid Build Coastguard Worker
755*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod19x0123456789ABCDEFlo = _mm256_mullo_epi16(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF);
756*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod19x0123456789ABCDEFhi = _mm256_srai_epi16(vprod19x0123456789ABCDEFlo, 15);
757*4bdc9457SAndroid Build Coastguard Worker
758*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod19x0123456789ABCDEFlo, vprod19x0123456789ABCDEFhi));
759*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod19x0123456789ABCDEFlo, vprod19x0123456789ABCDEFhi));
760*4bdc9457SAndroid Build Coastguard Worker
761*4bdc9457SAndroid Build Coastguard Worker const __m256i vi20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i20));
762*4bdc9457SAndroid Build Coastguard Worker const __m256i vk20x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 640)));
763*4bdc9457SAndroid Build Coastguard Worker i20 += 16;
764*4bdc9457SAndroid Build Coastguard Worker
765*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod20x0123456789ABCDEFlo = _mm256_mullo_epi16(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF);
766*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod20x0123456789ABCDEFhi = _mm256_srai_epi16(vprod20x0123456789ABCDEFlo, 15);
767*4bdc9457SAndroid Build Coastguard Worker
768*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod20x0123456789ABCDEFlo, vprod20x0123456789ABCDEFhi));
769*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod20x0123456789ABCDEFlo, vprod20x0123456789ABCDEFhi));
770*4bdc9457SAndroid Build Coastguard Worker
771*4bdc9457SAndroid Build Coastguard Worker const __m256i vi21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i21));
772*4bdc9457SAndroid Build Coastguard Worker const __m256i vk21x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 672)));
773*4bdc9457SAndroid Build Coastguard Worker i21 += 16;
774*4bdc9457SAndroid Build Coastguard Worker
775*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod21x0123456789ABCDEFlo = _mm256_mullo_epi16(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF);
776*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod21x0123456789ABCDEFhi = _mm256_srai_epi16(vprod21x0123456789ABCDEFlo, 15);
777*4bdc9457SAndroid Build Coastguard Worker
778*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod21x0123456789ABCDEFlo, vprod21x0123456789ABCDEFhi));
779*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod21x0123456789ABCDEFlo, vprod21x0123456789ABCDEFhi));
780*4bdc9457SAndroid Build Coastguard Worker
781*4bdc9457SAndroid Build Coastguard Worker const __m256i vi22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i22));
782*4bdc9457SAndroid Build Coastguard Worker const __m256i vk22x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 704)));
783*4bdc9457SAndroid Build Coastguard Worker i22 += 16;
784*4bdc9457SAndroid Build Coastguard Worker
785*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod22x0123456789ABCDEFlo = _mm256_mullo_epi16(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF);
786*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod22x0123456789ABCDEFhi = _mm256_srai_epi16(vprod22x0123456789ABCDEFlo, 15);
787*4bdc9457SAndroid Build Coastguard Worker
788*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod22x0123456789ABCDEFlo, vprod22x0123456789ABCDEFhi));
789*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod22x0123456789ABCDEFlo, vprod22x0123456789ABCDEFhi));
790*4bdc9457SAndroid Build Coastguard Worker
791*4bdc9457SAndroid Build Coastguard Worker const __m256i vi23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i23));
792*4bdc9457SAndroid Build Coastguard Worker const __m256i vk23x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 736)));
793*4bdc9457SAndroid Build Coastguard Worker i23 += 16;
794*4bdc9457SAndroid Build Coastguard Worker
795*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod23x0123456789ABCDEFlo = _mm256_mullo_epi16(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF);
796*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod23x0123456789ABCDEFhi = _mm256_srai_epi16(vprod23x0123456789ABCDEFlo, 15);
797*4bdc9457SAndroid Build Coastguard Worker
798*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod23x0123456789ABCDEFlo, vprod23x0123456789ABCDEFhi));
799*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod23x0123456789ABCDEFlo, vprod23x0123456789ABCDEFhi));
800*4bdc9457SAndroid Build Coastguard Worker
801*4bdc9457SAndroid Build Coastguard Worker const __m256i vi24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i24));
802*4bdc9457SAndroid Build Coastguard Worker const __m256i vk24x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 768)));
803*4bdc9457SAndroid Build Coastguard Worker i24 += 16;
804*4bdc9457SAndroid Build Coastguard Worker
805*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod24x0123456789ABCDEFlo = _mm256_mullo_epi16(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF);
806*4bdc9457SAndroid Build Coastguard Worker const __m256i vprod24x0123456789ABCDEFhi = _mm256_srai_epi16(vprod24x0123456789ABCDEFlo, 15);
807*4bdc9457SAndroid Build Coastguard Worker
808*4bdc9457SAndroid Build Coastguard Worker vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod24x0123456789ABCDEFlo, vprod24x0123456789ABCDEFhi));
809*4bdc9457SAndroid Build Coastguard Worker vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod24x0123456789ABCDEFlo, vprod24x0123456789ABCDEFhi));
810*4bdc9457SAndroid Build Coastguard Worker
811*4bdc9457SAndroid Build Coastguard Worker vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
812*4bdc9457SAndroid Build Coastguard Worker vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
813*4bdc9457SAndroid Build Coastguard Worker
814*4bdc9457SAndroid Build Coastguard Worker k += 16;
815*4bdc9457SAndroid Build Coastguard Worker
816*4bdc9457SAndroid Build Coastguard Worker __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
817*4bdc9457SAndroid Build Coastguard Worker __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
818*4bdc9457SAndroid Build Coastguard Worker
819*4bdc9457SAndroid Build Coastguard Worker const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t)));
820*4bdc9457SAndroid Build Coastguard Worker const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t) + 8 * sizeof(float)));
821*4bdc9457SAndroid Build Coastguard Worker vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
822*4bdc9457SAndroid Build Coastguard Worker vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
823*4bdc9457SAndroid Build Coastguard Worker
824*4bdc9457SAndroid Build Coastguard Worker const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
825*4bdc9457SAndroid Build Coastguard Worker vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
826*4bdc9457SAndroid Build Coastguard Worker vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
827*4bdc9457SAndroid Build Coastguard Worker
828*4bdc9457SAndroid Build Coastguard Worker vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
829*4bdc9457SAndroid Build Coastguard Worker vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
830*4bdc9457SAndroid Build Coastguard Worker
831*4bdc9457SAndroid Build Coastguard Worker w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
832*4bdc9457SAndroid Build Coastguard Worker
833*4bdc9457SAndroid Build Coastguard Worker const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
834*4bdc9457SAndroid Build Coastguard Worker __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
835*4bdc9457SAndroid Build Coastguard Worker __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), voutput_zero_point);
836*4bdc9457SAndroid Build Coastguard Worker
837*4bdc9457SAndroid Build Coastguard Worker const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
838*4bdc9457SAndroid Build Coastguard Worker
839*4bdc9457SAndroid Build Coastguard Worker __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
840*4bdc9457SAndroid Build Coastguard Worker vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
841*4bdc9457SAndroid Build Coastguard Worker
842*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(c >= 16) {
843*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
844*4bdc9457SAndroid Build Coastguard Worker output += 16;
845*4bdc9457SAndroid Build Coastguard Worker c -= 16;
846*4bdc9457SAndroid Build Coastguard Worker } else {
847*4bdc9457SAndroid Build Coastguard Worker if (c & 8) {
848*4bdc9457SAndroid Build Coastguard Worker _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
849*4bdc9457SAndroid Build Coastguard Worker vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
850*4bdc9457SAndroid Build Coastguard Worker output += 8;
851*4bdc9457SAndroid Build Coastguard Worker }
852*4bdc9457SAndroid Build Coastguard Worker if (c & 4) {
853*4bdc9457SAndroid Build Coastguard Worker unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
854*4bdc9457SAndroid Build Coastguard Worker vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
855*4bdc9457SAndroid Build Coastguard Worker output += 4;
856*4bdc9457SAndroid Build Coastguard Worker }
857*4bdc9457SAndroid Build Coastguard Worker if (c & 2) {
858*4bdc9457SAndroid Build Coastguard Worker unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
859*4bdc9457SAndroid Build Coastguard Worker vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
860*4bdc9457SAndroid Build Coastguard Worker output += 2;
861*4bdc9457SAndroid Build Coastguard Worker }
862*4bdc9457SAndroid Build Coastguard Worker if (c & 1) {
863*4bdc9457SAndroid Build Coastguard Worker *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
864*4bdc9457SAndroid Build Coastguard Worker output += 1;
865*4bdc9457SAndroid Build Coastguard Worker }
866*4bdc9457SAndroid Build Coastguard Worker c = 0;
867*4bdc9457SAndroid Build Coastguard Worker }
868*4bdc9457SAndroid Build Coastguard Worker } while (c != 0);
869*4bdc9457SAndroid Build Coastguard Worker }
870*4bdc9457SAndroid Build Coastguard Worker
871*4bdc9457SAndroid Build Coastguard Worker output = (int8_t*) ((uintptr_t) output + output_increment);
872*4bdc9457SAndroid Build Coastguard Worker } while (--output_width != 0);
873*4bdc9457SAndroid Build Coastguard Worker }
874