1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx512skx-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/intrinsics-polyfill.h>
16
17
xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx512skx_mul32(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx512skx_mul32(
19 size_t channels,
20 size_t output_width,
21 const uint8_t** input,
22 const void* weights,
23 uint8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_MSAN
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m512 vscale = _mm512_load_ps(params->fp32_avx512.scale);
34 const __m512 voutput_max_less_zero_point = _mm512_load_ps(params->fp32_avx512.output_max_less_zero_point);
35 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx512.output_zero_point);
36 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx512.output_min);
37
38 const __m512i vk_zero_point = _mm512_cvtepu16_epi32(_mm256_load_si256((const __m256i*) params->fp32_avx512.kernel_zero_point));
39 do {
40 const uint8_t* i0 = input[0];
41 assert(i0 != NULL);
42 if XNN_UNPREDICTABLE(i0 != zero) {
43 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
44 }
45 const uint8_t* i1 = input[1];
46 assert(i1 != NULL);
47 if XNN_UNPREDICTABLE(i1 != zero) {
48 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
49 }
50 const uint8_t* i2 = input[2];
51 assert(i2 != NULL);
52 if XNN_UNPREDICTABLE(i2 != zero) {
53 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
54 }
55 const uint8_t* i3 = input[3];
56 assert(i3 != NULL);
57 if XNN_UNPREDICTABLE(i3 != zero) {
58 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
59 }
60 const uint8_t* i4 = input[4];
61 assert(i4 != NULL);
62 if XNN_UNPREDICTABLE(i4 != zero) {
63 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
64 }
65 const uint8_t* i5 = input[5];
66 assert(i5 != NULL);
67 if XNN_UNPREDICTABLE(i5 != zero) {
68 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
69 }
70 const uint8_t* i6 = input[6];
71 assert(i6 != NULL);
72 if XNN_UNPREDICTABLE(i6 != zero) {
73 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
74 }
75 const uint8_t* i7 = input[7];
76 assert(i7 != NULL);
77 if XNN_UNPREDICTABLE(i7 != zero) {
78 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
79 }
80 const uint8_t* i8 = input[8];
81 assert(i8 != NULL);
82 if XNN_UNPREDICTABLE(i8 != zero) {
83 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
84 }
85 input = (const uint8_t**) ((uintptr_t) input + input_stride);
86
87 size_t c = channels;
88 const void* w = weights;
89 for (; c >= 16; c -= 16) {
90 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
91
92
93 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i0));
94 const __m512i vk0x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
95 i0 += 16;
96
97 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
98
99 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i1));
100 const __m512i vk1x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
101 i1 += 16;
102
103 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
104
105 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i2));
106 const __m512i vk2x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
107 i2 += 16;
108
109 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
110
111 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i3));
112 const __m512i vk3x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
113 i3 += 16;
114
115 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
116
117 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i4));
118 const __m512i vk4x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
119 i4 += 16;
120
121 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
122
123 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i5));
124 const __m512i vk5x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
125 i5 += 16;
126
127 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
128
129 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i6));
130 const __m512i vk6x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
131 i6 += 16;
132
133 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
134
135 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i7));
136 const __m512i vk7x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
137 i7 += 16;
138
139 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
140
141 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8));
142 const __m512i vk8x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
143 i8 += 16;
144
145 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
146
147 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t));
148
149 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
150
151 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
152
153 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
154
155 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
156
157 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
158
159 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
160 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
161 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
162
163 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
164
165 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
166 output += 16;
167 }
168 if XNN_UNLIKELY(c != 0) {
169 // Prepare mask for valid 8-bit elements (depends on nc).
170 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1)));
171 {
172 __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
173
174
175 const __m512i vi0x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i0));
176 const __m512i vk0x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
177
178 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
179
180 const __m512i vi1x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i1));
181 const __m512i vk1x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
182
183 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
184
185 const __m512i vi2x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i2));
186 const __m512i vk2x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
187
188 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
189
190 const __m512i vi3x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i3));
191 const __m512i vk3x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
192
193 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
194
195 const __m512i vi4x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i4));
196 const __m512i vk4x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
197
198 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
199
200 const __m512i vi5x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i5));
201 const __m512i vk5x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
202
203 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
204
205 const __m512i vi6x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i6));
206 const __m512i vk6x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
207
208 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
209
210 const __m512i vi7x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i7));
211 const __m512i vk7x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
212
213 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
214
215 const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8));
216 const __m512i vk8x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
217
218 vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
219
220
221 __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
222 vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale);
223 vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
224 vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
225
226
227 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), voutput_zero_point);
228
229 const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
230 const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
231 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
232 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
233
234 _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
235 output = (uint8_t*) ((uintptr_t) output + c);
236 }
237 }
238
239 output = (uint8_t*) ((uintptr_t) output + output_increment);
240 } while (--output_width != 0);
241 }
242