1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c2-neon-mull-dup.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/gemm.h>
15 #include <xnnpack/math.h>
16
xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r(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)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 2);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(int8_t) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 kc = round_up_po2(kc, 2 * sizeof(int8_t));
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
56
57 while (k >= 8 * sizeof(int8_t)) {
58 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
59 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
60 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
61 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
62 const int16x4_t va10 = vld1_dup_s16((const void*)a1);
63 const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
64 const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
65 const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
66
67 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75
76 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
77 const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
78
79 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
80 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
81 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
82 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
83 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
84 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
85 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
86 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
87 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
88 const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
89
90 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
91 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
92 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
93 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
94 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
95 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
96 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
97 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
98 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
99 const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
100
101 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
102 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
103 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
104 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
105 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
106 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
107 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
108 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
109 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
110 const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
111
112 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
113 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
114 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
115 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
116 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
117 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
118 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
119 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
120
121 k -= 8 * sizeof(int8_t);
122 }
123
124 if XNN_UNLIKELY(k != 0) {
125 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
126 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
127
128 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130
131 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
132 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
133 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
134 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
135 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
136 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
137 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
138 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
139 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
140 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
141
142 if (k > 2 * sizeof(int8_t)) {
143 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
145
146 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
147 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
148 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
149 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
150 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
151 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
152 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
153 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
154 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
155 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
156
157 if (k > 4 * sizeof(int8_t)) {
158 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160
161 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
162 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
163 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
164 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
165 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
166 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
167 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
168 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
169 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
170 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
171 }
172 }
173 }
174
175 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
176 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
177 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
178
179 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
180 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
181 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
182 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
183
184 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
185 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
186 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
187 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
188
189 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
190 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
191 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
192 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
193
194 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
195 #if XNN_ARCH_ARM64
196 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
197 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
198
199 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
200 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
201
202 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
203 #else
204 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
205 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
206
207 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
208 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
209
210 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
211 #endif
212
213 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
214 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
215
216 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
217 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
218
219 if (nc >= 8) {
220 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
221 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
222
223 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
224 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
225
226 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
227 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
228
229 nc -= 8;
230 } else {
231 // Final case where not all of the 8 columns fit in the destination.
232 if (nc & 4) {
233 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
234 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
235 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
236 }
237 if (nc & 2) {
238 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
239 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
240 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
241 }
242 if (nc & 1) {
243 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
244 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
245 }
246
247 nc = 0;
248 }
249 } while (nc != 0);
250 }
251