1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-neon-mull-shuffle.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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal(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)])19 void xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 2);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (2 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 kc = round_up_po2(kc, 8 * sizeof(int8_t));
51 do {
52 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
54 int32x4_t vacc1x0123 = vacc0x0123;
55 int32x4_t vacc1x4567 = vacc0x4567;
56
57 size_t p = ks;
58 do {
59 const int8_t* restrict a0 = a[0];
60 if XNN_UNPREDICTABLE(a0 != zero) {
61 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
62 }
63 const int8_t* restrict a1 = a[1];
64 if XNN_UNPREDICTABLE(a1 != zero) {
65 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
66 }
67 a += 2;
68
69 size_t k = kc;
70 while (k >= 16 * sizeof(int8_t)) {
71 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
72 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
73 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
74 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
75
76 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84
85 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
86 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
87 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
88 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
89 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
90 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
91 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
92 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
93 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
94 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
95 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
96 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
97 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
98 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
99 va0x0 = vext_s8(va0x0, va0x0, 2);
100 va0x1 = vext_s8(va0x1, va0x1, 2);
101 va1x0 = vext_s8(va1x0, va1x0, 2);
102 va1x1 = vext_s8(va1x1, va1x1, 2);
103 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
104 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
105 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
106 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
107 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
108 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
109 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
110 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
111 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
112 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
113 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
114 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
115 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
116 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
117 va0x0 = vext_s8(va0x0, va0x0, 2);
118 va0x1 = vext_s8(va0x1, va0x1, 2);
119 va1x0 = vext_s8(va1x0, va1x0, 2);
120 va1x1 = vext_s8(va1x1, va1x1, 2);
121 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
122 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
123 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
124 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
125 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
126 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
127 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
128 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
129 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
130 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
131 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
132 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
133 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
134 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
135 va0x0 = vext_s8(va0x0, va0x0, 2);
136 va0x1 = vext_s8(va0x1, va0x1, 2);
137 va1x0 = vext_s8(va1x0, va1x0, 2);
138 va1x1 = vext_s8(va1x1, va1x1, 2);
139 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
140 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
141 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
142 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
143 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
144 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
145 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
146 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
147 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
148 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
149 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
150 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
151 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
152 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
153
154 k -= 16 * sizeof(int8_t);
155 }
156 if (k != 0) {
157 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
158 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
159
160 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
162 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
163 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
164 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
165 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
166 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
167 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
168
169 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
170 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
171 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
172 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
173 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
174 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
175 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
176 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
177 va0x0 = vext_s8(va0x0, va0x0, 2);
178 va1x0 = vext_s8(va1x0, va1x0, 2);
179 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
180 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
181 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
182 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
183 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
184 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
185 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
186 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
187 va0x0 = vext_s8(va0x0, va0x0, 2);
188 va1x0 = vext_s8(va1x0, va1x0, 2);
189 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
190 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
191 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
192 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
193 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
194 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
195 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
196 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
197 va0x0 = vext_s8(va0x0, va0x0, 2);
198 va1x0 = vext_s8(va1x0, va1x0, 2);
199 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
200 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
201 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
202 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
203 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
204 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
205 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
206 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
207
208 }
209
210 p -= 2 * sizeof(void*);
211 } while (p != 0);
212
213 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
214 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
215 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
216 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
217
218 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
219 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
220 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
221 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
222 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
223
224 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
225 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
226 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
227 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
228
229 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
230 #if XNN_ARCH_ARM64
231 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
232 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
233
234 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
235 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
236
237 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
238 #else
239 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
240 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
241
242 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
243 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
244
245 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
246 #endif
247
248 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
249 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
250
251 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max);
252 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
253
254 if (nc >= 8) {
255 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
256 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
257
258 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
259 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
260
261 a = (const int8_t**restrict) ((uintptr_t) a - ks);
262
263 nc -= 8;
264 } else {
265 if (nc & 4) {
266 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
267 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
268 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
269 }
270 if (nc & 2) {
271 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
272 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
273 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
274 }
275 if (nc & 1) {
276 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
277 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
278 }
279
280 nc = 0;
281 }
282 } while (nc != 0);
283 }
284