1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx2-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__avx2_mul32(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 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__avx2_mul32(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 input = (const int8_t**) ((uintptr_t) input + input_stride);
50
51 size_t c = channels;
52 const void* w = weights;
53 for (; c >= 16; c -= 16) {
54 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
55 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 8));
56
57
58 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
59 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
60 const __m256i vi0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
61 const __m256i vk0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t))));
62 i0 += 16;
63
64 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
65 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi0x89ABCDEF, vk0x89ABCDEF));
66
67 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
68 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
69 const __m256i vi1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
70 const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t))));
71 i1 += 16;
72
73 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
74 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF));
75
76 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
77 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
78 const __m256i vi2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
79 const __m256i vk2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t))));
80 i2 += 16;
81
82 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
83 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF));
84
85 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t));
86
87 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
88 __m256 vscaled89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
89
90 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w);
91 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) w + 8);
92 w = (const void*) ((const float*) w + 16);
93 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale01234567);
94 vscaled89ABCDEF = _mm256_mul_ps(vscaled89ABCDEF, vscale89ABCDEF);
95
96 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
97 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
98 vscaled89ABCDEF = _mm256_min_ps(vscaled89ABCDEF, voutput_max_less_zero_point);
99
100 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
101 vacc89ABCDEF = _mm256_cvtps_epi32(vscaled89ABCDEF);
102
103 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
104 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
105
106 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
107
108 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
109 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
110
111 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
112 output += 16;
113 }
114 if XNN_UNLIKELY(c != 0) {
115 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
116 do {
117 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
118
119
120 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
121 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) k));
122 i0 += 8;
123
124 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
125
126 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
127 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 16)));
128 i1 += 8;
129
130 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
131
132 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
133 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 32)));
134 i2 += 8;
135
136 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
137
138 k += 8;
139
140 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
141 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
142 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale01234567);
143 vscaled01234567 = _mm256_min_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point));
144 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
145
146 w = (const void*) ((const int32_t*) w + 8);
147
148 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
149 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
150
151 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
152
153 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
154 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
155
156 if XNN_LIKELY(c >= 8) {
157 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
158 output += 8;
159 c -= 8;
160 } else {
161 if (c & 4) {
162 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
163 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
164 output += 4;
165 }
166 if (c & 2) {
167 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
168 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
169 output += 2;
170 }
171 if (c & 1) {
172 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
173 output += 1;
174 }
175 c = 0;
176 }
177 } while (c != 0);
178 }
179
180 output = (int8_t*) ((uintptr_t) output + output_increment);
181 } while (--output_width != 0);
182 }
183