1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const int8_t** restrict a,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const int8_t* zero,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (1 * sizeof(void*)) == 0);
37 assert(a_offset % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 2 * sizeof(int8_t));
43 int8_t* c0 = c;
44
45 do {
46 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
48
49 size_t p = ks;
50 do {
51 const int8_t* restrict a0 = a[0];
52 if XNN_UNPREDICTABLE(a0 != zero) {
53 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
54 }
55 a += 1;
56
57 size_t k = kc;
58
59 while (k >= 16 * sizeof(int8_t)) {
60 const int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
61 const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
62 const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
63 const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
64 const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
65 const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
66 const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
67 const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
68
69 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77
78 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
79 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
80
81 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
82 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
84 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
85 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
86 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
88 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
89 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
90 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
91
92 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
93 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
95 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
96 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
97 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
99 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
100 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
101 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
102
103 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
104 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
106 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
107 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
108 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
110 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
111 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
112 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
113
114 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
115 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
117 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
118 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
119 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
121 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
122
123 k -= 16 * sizeof(int8_t);
124 }
125
126 if (k >= 8 * sizeof(int8_t)) {
127 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
128 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
129 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
130 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
131
132 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140
141 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
142
143 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
144 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
145 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
146 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
147 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
148
149 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
150 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
151 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
152 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
153 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
154
155 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
156 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
157 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
158 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
159 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
160
161 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
162 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
163 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
164 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
165
166 k -= 8 * sizeof(int8_t);
167 }
168
169 if XNN_UNLIKELY(k != 0) {
170 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
171
172 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174
175 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
176 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
177 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
178 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
179 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
180
181 if (k > 2 * sizeof(int8_t)) {
182 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184
185 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
186 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
187 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
188 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
189 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
190
191 if (k > 4 * sizeof(int8_t)) {
192 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
193 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
194
195 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
196 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
197 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
198 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
199 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
200 }
201 }
202 }
203 p -= 1 * sizeof(void*);
204 } while (p != 0);
205
206 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
207 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
208 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
209
210 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
211 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
212
213 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
214 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
215
216 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
217 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
218
219 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
220 #if XNN_ARCH_ARM64
221 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
222
223 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
224
225 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
226 #else
227 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
228
229 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
230
231 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
232 #endif
233
234 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
235 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
236
237 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
238 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
239
240 if (nc >= 8) {
241 vst1_s8(c0 + 0, vout0x01234567);
242
243 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
244
245 a = (const int8_t**restrict) ((uintptr_t) a - ks);
246
247 nc -= 8;
248 } else {
249 if (nc & 4) {
250 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
251 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
252 }
253 if (nc & 2) {
254 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
255 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
256 }
257 if (nc & 1) {
258 vst1_lane_s8(c0, vout0x01234567, 0);
259 }
260
261 nc = 0;
262 }
263 } while (nc != 0);
264 }
265