1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/neon-mlal-lane.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_1x16__neon_mlal_lane(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_1x16__neon_mlal_lane(
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 <= 1);
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
42 do {
43 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
44 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
45 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
46 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
47
48 size_t k = kc;
49 while (k >= 8 * sizeof(int8_t)) {
50 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
51 const int16x8_t vxa0 = vmovl_s8(va0);
52
53 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
54 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
55
56 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
57 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
58 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
59 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
60
61 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
62 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
63 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
64 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
65
66 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
67 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
68 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
69 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
70
71 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
72 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
73 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
74 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
75
76 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
77 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
78 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
79 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
80
81 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
82 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
83 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
84 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
85
86 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
87 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
88 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
89 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
90
91 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
92 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
93
94
95 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
96 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
97
98 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
99 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
100 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
101 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
102
103 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
104 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
105 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
106 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
107
108 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
109 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
110 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
111 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
112
113 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
114 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
115 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
116 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
117
118 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
119 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
120 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
121 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
122
123 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
124 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
125 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
126 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
127
128 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
129 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
130 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
131 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
132
133 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
134 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
135
136 k -= 8 * sizeof(int8_t);
137 }
138 if XNN_UNLIKELY(k != 0) {
139 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
140 const int16x8_t vxa0 = vmovl_s8(va0);
141
142 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
143 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
144 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
145 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
146
147 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
148 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
149 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
150 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
151
152 if (k >= 2 * sizeof(int8_t)) {
153 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
154 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
155 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
156 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
157
158 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
159 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
160 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
161 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
162
163 if (k > 2 * sizeof(int8_t)) {
164 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
165 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
166 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
167 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
168
169 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
170 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
171 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
172 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
173
174 if (k >= 4 * sizeof(int8_t)) {
175 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
176 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
177 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
178 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
179
180 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
181 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
182 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
183 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
184
185 if (k > 4 * sizeof(int8_t)) {
186 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
187 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
188 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
189 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
190
191 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
192 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
193 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
194 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
195
196 if (k >= 6 * sizeof(int8_t)) {
197 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
198 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
199 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
200 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
201
202 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
203 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
204 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
205 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
206
207 if (k > 6 * sizeof(int8_t)) {
208 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
209 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
210 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
211 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
212
213 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
214 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
215 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
216 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
217 }
218 }
219 }
220 }
221 }
222 }
223 }
224
225 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
226 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
227 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
228
229 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
230 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
231 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
232 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
233
234 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
235 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
236 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
237 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
238
239 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
240 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
241 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
242 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
243
244 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
245 #if XNN_ARCH_ARM64
246 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
247 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
248
249 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
250 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
251
252 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
253 #else
254 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
255 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
256
257 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
258 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
259
260 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
261 #endif
262
263 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
264 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
265
266 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
267 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
268
269 if (nc >= 16) {
270 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
271
272 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
273
274 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
275
276 nc -= 16;
277 } else {
278 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
279 if (nc & 8) {
280 vst1_s8(c0, vout0x01234567); c0 += 8;
281 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
282 }
283 if (nc & 4) {
284 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
285 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
286 }
287 if (nc & 2) {
288 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
289 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
290 }
291 if (nc & 1) {
292 vst1_lane_s8(c0, vout0x01234567, 0);
293 }
294
295 nc = 0;
296 }
297 } while (nc != 0);
298 }
299