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_1x8c2__neon_mlal_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)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup(
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 <= 1);
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
42 do {
43 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
44 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45
46 size_t k = kc;
47
48 while (k >= 16 * sizeof(int8_t)) {
49 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
50 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
51 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
52 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
53 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
54 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
55 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59
60 const int8x8_t va0c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0));
61 const int8x8_t va0c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0));
62
63 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
64 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
66 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
67 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
68 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
70 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
71 const int8x8_t va0c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1));
72 const int8x8_t va0c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1));
73
74 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
75 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
77 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
78 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
79 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
81 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
82 const int8x8_t va0c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2));
83 const int8x8_t va0c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2));
84
85 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
86 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
88 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
89 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
90 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
92 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
93 const int8x8_t va0c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3));
94 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3));
95
96 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
97 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
99 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
100 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
101 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
103 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
104
105 k -= 16 * sizeof(int8_t);
106 }
107
108 if (k >= 8 * sizeof(int8_t)) {
109 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
110
111 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119
120 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
121
122 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
123 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
124 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
125 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
126 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
127
128 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
129 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
130 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
131 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
132 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
133
134 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
135 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
136 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
137 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
138 const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
139
140 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
141 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
142 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
143 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
144
145 k -= 8 * sizeof(int8_t);
146 }
147
148 if XNN_UNLIKELY(k != 0) {
149 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
150
151 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153
154 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
155 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
156 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
157 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
158 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
159
160 if (k > 2 * sizeof(int8_t)) {
161 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163
164 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
165 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
166 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
167 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
168 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
169
170 if (k > 4 * sizeof(int8_t)) {
171 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173
174 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
175 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
176 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
177 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
178 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
179 }
180 }
181 }
182
183 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
184 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
185 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
186
187 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
188 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
189
190 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
191 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
192
193 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
194 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
195
196 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
197 #if XNN_ARCH_ARM64
198 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
199
200 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
201
202 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
203 #else
204 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
205
206 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
207
208 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
209 #endif
210
211 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
212 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
213
214 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
215 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
216
217 if (nc >= 8) {
218 vst1_s8(c0 + 0, vout0x01234567);
219
220 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
221
222 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
223
224 nc -= 8;
225 } else {
226 // Final case where not all of the 8 columns fit in the destination.
227 if (nc & 4) {
228 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
229 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
230 }
231 if (nc & 2) {
232 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
233 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
234 }
235 if (nc & 1) {
236 vst1_lane_s8(c0, vout0x01234567, 0);
237 }
238
239 nc = 0;
240 }
241 } while (nc != 0);
242 }
243