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 #include <xnnpack/unaligned.h>
22
23
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)])24 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32(
25 size_t channels,
26 size_t output_width,
27 const uint8_t** input,
28 const void* weights,
29 uint8_t* output,
30 size_t input_stride,
31 size_t output_increment,
32 size_t input_offset,
33 const uint8_t* zero,
34 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(channels != 0);
37 assert(output_width != 0);
38
39 const __m128i vk_zero_point = _mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) params->fp32_sse2.kernel_zero_point));
40 do {
41 const uint8_t* i0 = input[0];
42 assert(i0 != NULL);
43 if XNN_UNPREDICTABLE(i0 != zero) {
44 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
45 }
46 const uint8_t* i1 = input[1];
47 assert(i1 != NULL);
48 if XNN_UNPREDICTABLE(i1 != zero) {
49 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
50 }
51 const uint8_t* i2 = input[2];
52 assert(i2 != NULL);
53 if XNN_UNPREDICTABLE(i2 != zero) {
54 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
55 }
56 const uint8_t* i3 = input[3];
57 assert(i3 != NULL);
58 if XNN_UNPREDICTABLE(i3 != zero) {
59 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
60 }
61 const uint8_t* i4 = input[4];
62 assert(i4 != NULL);
63 if XNN_UNPREDICTABLE(i4 != zero) {
64 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
65 }
66 const uint8_t* i5 = input[5];
67 assert(i5 != NULL);
68 if XNN_UNPREDICTABLE(i5 != zero) {
69 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
70 }
71 const uint8_t* i6 = input[6];
72 assert(i6 != NULL);
73 if XNN_UNPREDICTABLE(i6 != zero) {
74 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
75 }
76 const uint8_t* i7 = input[7];
77 assert(i7 != NULL);
78 if XNN_UNPREDICTABLE(i7 != zero) {
79 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
80 }
81 const uint8_t* i8 = input[8];
82 assert(i8 != NULL);
83 if XNN_UNPREDICTABLE(i8 != zero) {
84 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
85 }
86 input = (const uint8_t**) ((uintptr_t) input + input_stride);
87
88 size_t c = channels;
89 const void* w = weights;
90 for (; c >= 8; c -= 8) {
91 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
92 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
93
94
95 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
96 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t))))), vk_zero_point);
97 const __m128i vi0x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
98 const __m128i vk0x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(uint8_t))))), vk_zero_point);
99 i0 += 8;
100
101 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
102 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
103
104 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
105 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t))))), vk_zero_point);
106 const __m128i vi1x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
107 const __m128i vk1x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(uint8_t))))), vk_zero_point);
108 i1 += 8;
109
110 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
111 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
112
113 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
114 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t))))), vk_zero_point);
115 const __m128i vi2x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
116 const __m128i vk2x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(uint8_t))))), vk_zero_point);
117 i2 += 8;
118
119 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
120 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
121
122 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
123 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t))))), vk_zero_point);
124 const __m128i vi3x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
125 const __m128i vk3x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(uint8_t))))), vk_zero_point);
126 i3 += 8;
127
128 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
129 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
130
131 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
132 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t))))), vk_zero_point);
133 const __m128i vi4x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
134 const __m128i vk4x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(uint8_t))))), vk_zero_point);
135 i4 += 8;
136
137 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
138 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
139
140 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
141 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t))))), vk_zero_point);
142 const __m128i vi5x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
143 const __m128i vk5x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(uint8_t))))), vk_zero_point);
144 i5 += 8;
145
146 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
147 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
148
149 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
150 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t))))), vk_zero_point);
151 const __m128i vi6x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
152 const __m128i vk6x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 52 * sizeof(uint8_t))))), vk_zero_point);
153 i6 += 8;
154
155 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
156 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
157
158 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
159 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t))))), vk_zero_point);
160 const __m128i vi7x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
161 const __m128i vk7x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 60 * sizeof(uint8_t))))), vk_zero_point);
162 i7 += 8;
163
164 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
165 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
166
167 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
168 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t))))), vk_zero_point);
169 const __m128i vi8x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
170 const __m128i vk8x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 68 * sizeof(uint8_t))))), vk_zero_point);
171 i8 += 8;
172
173 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
174 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
175
176 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t));
177
178 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
179 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
180
181 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
182 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
183 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
184
185 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
186 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
187 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
188
189 vacc0123 = _mm_cvtps_epi32(vscaled0123);
190 vacc4567 = _mm_cvtps_epi32(vscaled4567);
191
192 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
193 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
194
195 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
196 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
197 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
198
199 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
200 output += 8;
201 }
202 if XNN_UNLIKELY(c != 0) {
203 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 8);
204 do {
205 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
206
207 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
208 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) k))), vk_zero_point);
209 i0 += 4;
210
211 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
212 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
213 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 8)))), vk_zero_point);
214 i1 += 4;
215
216 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
217 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
218 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16)))), vk_zero_point);
219 i2 += 4;
220
221 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
222 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
223 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 24)))), vk_zero_point);
224 i3 += 4;
225
226 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
227 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
228 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32)))), vk_zero_point);
229 i4 += 4;
230
231 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
232 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
233 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 40)))), vk_zero_point);
234 i5 += 4;
235
236 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
237 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
238 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48)))), vk_zero_point);
239 i6 += 4;
240
241 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
242 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
243 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 56)))), vk_zero_point);
244 i7 += 4;
245
246 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
247 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
248 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64)))), vk_zero_point);
249 i8 += 4;
250
251 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
252
253 k += 4;
254
255 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
256 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.scale));
257 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.output_max_less_zero_point));
258 vacc0123 = _mm_cvtps_epi32(vscaled0123);
259
260 w = (const void*) ((const int32_t*) w + 4);
261
262 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
263 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
264
265 vout0123 = _mm_packus_epi16(vout0123, vout0123);
266 vout0123 = _mm_max_epu8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
267
268 if XNN_LIKELY(c >= 4) {
269 _mm_storeu_si32(output, vout0123);
270 output += 4;
271 c -= 4;
272 } else {
273 if (c & 2) {
274 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
275 vout0123 = _mm_srli_epi32(vout0123, 16);
276 output += 2;
277 }
278 if (c & 1) {
279 *output = (uint8_t) _mm_extract_epi8(vout0123, 0);
280 output += 1;
281 }
282 c = 0;
283 }
284 } while (c != 0);
285 }
286
287 output = (uint8_t*) ((uintptr_t) output + output_increment);
288 } while (--output_width != 0);
289 }
290