1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/dwconv.h>
20 #include <xnnpack/intrinsics-polyfill.h>
21
22
xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__xop_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)])23 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32(
24 size_t channels,
25 size_t output_width,
26 const uint8_t** input,
27 const void* weights,
28 uint8_t* output,
29 size_t input_stride,
30 size_t output_increment,
31 size_t input_offset,
32 const uint8_t* zero,
33 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37
38 const __m128i vk_zero_point = _mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) params->fp32_sse2.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 >= 8; c -= 8) {
90 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
91 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
92
93
94 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i0));
95 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
96 const __m128i vi0x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i0 + 4));
97 const __m128i vk0x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(uint8_t)))), vk_zero_point);
98 i0 += 8;
99
100 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
101 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
102
103 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i1));
104 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)))), vk_zero_point);
105 const __m128i vi1x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i1 + 4));
106 const __m128i vk1x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(uint8_t)))), vk_zero_point);
107 i1 += 8;
108
109 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
110 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
111
112 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i2));
113 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
114 const __m128i vi2x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i2 + 4));
115 const __m128i vk2x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(uint8_t)))), vk_zero_point);
116 i2 += 8;
117
118 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
119 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
120
121 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i3));
122 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)))), vk_zero_point);
123 const __m128i vi3x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i3 + 4));
124 const __m128i vk3x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(uint8_t)))), vk_zero_point);
125 i3 += 8;
126
127 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
128 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
129
130 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i4));
131 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
132 const __m128i vi4x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i4 + 4));
133 const __m128i vk4x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(uint8_t)))), vk_zero_point);
134 i4 += 8;
135
136 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
137 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
138
139 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i5));
140 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)))), vk_zero_point);
141 const __m128i vi5x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i5 + 4));
142 const __m128i vk5x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(uint8_t)))), vk_zero_point);
143 i5 += 8;
144
145 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
146 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
147
148 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i6));
149 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
150 const __m128i vi6x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i6 + 4));
151 const __m128i vk6x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 52 * sizeof(uint8_t)))), vk_zero_point);
152 i6 += 8;
153
154 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
155 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
156
157 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i7));
158 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)))), vk_zero_point);
159 const __m128i vi7x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i7 + 4));
160 const __m128i vk7x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 60 * sizeof(uint8_t)))), vk_zero_point);
161 i7 += 8;
162
163 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
164 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
165
166 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i8));
167 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
168 const __m128i vi8x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i8 + 4));
169 const __m128i vk8x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 68 * sizeof(uint8_t)))), vk_zero_point);
170 i8 += 8;
171
172 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
173 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
174
175 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t));
176
177 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
178 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
179
180 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
181 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
182 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
183
184 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
185 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
186 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
187
188 vacc0123 = _mm_cvtps_epi32(vscaled0123);
189 vacc4567 = _mm_cvtps_epi32(vscaled4567);
190
191 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
192 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
193
194 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
195 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
196 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
197
198 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
199 output += 8;
200 }
201 if XNN_UNLIKELY(c != 0) {
202 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 8);
203 do {
204 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
205
206 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i0));
207 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32(k)), vk_zero_point);
208 i0 += 4;
209
210 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
211 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i1));
212 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 8))), vk_zero_point);
213 i1 += 4;
214
215 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
216 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i2));
217 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 16))), vk_zero_point);
218 i2 += 4;
219
220 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
221 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i3));
222 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 24))), vk_zero_point);
223 i3 += 4;
224
225 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
226 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i4));
227 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 32))), vk_zero_point);
228 i4 += 4;
229
230 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
231 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i5));
232 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 40))), vk_zero_point);
233 i5 += 4;
234
235 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
236 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i6));
237 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 48))), vk_zero_point);
238 i6 += 4;
239
240 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
241 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i7));
242 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 56))), vk_zero_point);
243 i7 += 4;
244
245 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
246 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i8));
247 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 64))), vk_zero_point);
248 i8 += 4;
249
250 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
251
252 k += 4;
253
254 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
255 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.scale));
256 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.output_max_less_zero_point));
257 vacc0123 = _mm_cvtps_epi32(vscaled0123);
258
259 w = (const void*) ((const int32_t*) w + 4);
260
261 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
262 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
263
264 vout0123 = _mm_packus_epi16(vout0123, vout0123);
265 vout0123 = _mm_max_epu8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
266
267 if XNN_LIKELY(c >= 4) {
268 _mm_storeu_si32(output, vout0123);
269 output += 4;
270 c -= 4;
271 } else {
272 if (c & 2) {
273 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
274 vout0123 = _mm_srli_epi32(vout0123, 16);
275 output += 2;
276 }
277 if (c & 1) {
278 *output = (uint8_t) _mm_extract_epi8(vout0123, 0);
279 output += 1;
280 }
281 c = 0;
282 }
283 } while (c != 0);
284 }
285
286 output = (uint8_t*) ((uintptr_t) output + output_increment);
287 } while (--output_width != 0);
288 }
289