1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/neon-mull-addw-dup.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 <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 do {
49 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51 int32x4_t vacc1x0123 = vacc0x0123;
52 int32x4_t vacc1x4567 = vacc0x4567;
53
54 size_t k = kc;
55 while (k >= 8 * sizeof(int8_t)) {
56 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
57 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
58
59 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60
61 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
62 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
63 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
64 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
65 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
66 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
67 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68
69 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
70 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
71 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
72 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
73 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
74 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
75 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76
77 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
78 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
79 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
80 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
81 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
82 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
83 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84
85 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
86 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
87 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
88 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
89 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
90 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
91 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92
93 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
94 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
95 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
96 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
97 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
98 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
99 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100
101 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
102 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
103 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
104 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
105 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
106 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
107 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108
109 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
110 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
111 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
112 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
113 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
114 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
115 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116
117 const int16x8_t vprod0x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va0, 7));
118 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c7));
119 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c7));
120 const int16x8_t vprod1x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va1, 7));
121 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c7));
122 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c7));
123
124 k -= 8 * sizeof(int8_t);
125 }
126 if XNN_UNLIKELY(k != 0) {
127 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
128 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
129
130 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131
132 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
133 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
134 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
135 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
136 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
137 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
138
139 if (k >= 2 * sizeof(int8_t)) {
140 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141
142 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
143 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
144 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
145 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
146 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
147 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
148
149 if (k > 2 * sizeof(int8_t)) {
150 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151
152 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
153 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
154 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
155 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
156 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
157 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
158
159 if (k >= 4 * sizeof(int8_t)) {
160 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161
162 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
163 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
164 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
165 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
166 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
167 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
168
169 if (k > 4 * sizeof(int8_t)) {
170 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171
172 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
173 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
174 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
175 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
176 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
177 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
178
179 if (k >= 6 * sizeof(int8_t)) {
180 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181
182 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
183 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
184 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
185 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
186 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
187 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
188
189 if (k > 6 * sizeof(int8_t)) {
190 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
191
192 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
193 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
194 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
195 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
196 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
197 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
198 }
199 }
200 }
201 }
202 }
203 }
204 }
205
206 // Post-accumulation work
207 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
208 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
209 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
210
211 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
212 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
213 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
214 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
215
216 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
217 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
218 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
219 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
220
221 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
222 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
223 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
224 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
225
226 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
227 #if XNN_ARCH_ARM64
228 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
229 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
230
231 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
232 #else
233 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
234 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
235
236 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
237 #endif
238 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
239 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
240
241 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
242
243 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
244
245 if (nc >= 8) {
246 // Main case where there the 8 columns fit in the destination.
247 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
248 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
249
250 // Advance to the next 8 columns.
251 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
252 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
253
254 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
255 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
256
257 nc -= 8;
258 } else {
259 // Final case where not all of the 8 columns fit in the destination.
260 if (nc & 4) {
261 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
262 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
263 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
264 }
265 if (nc & 2) {
266 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
267 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
268 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
269 }
270 if (nc & 1) {
271 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
272 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
273 }
274
275 nc = 0;
276 }
277 } while (nc != 0);
278 }
279