1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-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_qc8_igemm_minmax_fp32_ukernel_2x8c4s2__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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4s2__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_qc8_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 vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
53 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
54 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
55 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
56 int32x4_t vacc1x01 = vacc0x01;
57 int32x4_t vacc1x23 = vacc0x23;
58 int32x4_t vacc1x45 = vacc0x45;
59 int32x4_t vacc1x67 = vacc0x67;
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const int8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 a += 2;
72
73 size_t k = kc;
74 while (k >= 16 * sizeof(int8_t)) {
75 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
76 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
77 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
78 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
79
80 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88
89 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
90 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
91 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
92 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
93 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
94 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
95 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
96 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
97 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
98 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
99 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
100 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
101 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
102 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
103 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
104 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
105 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
106 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
107 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
108 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
109 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
110 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
111 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
112 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
113 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
114 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
115 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
116 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
117 va0x0 = vext_s8(va0x0, va0x0, 4);
118 va0x1 = vext_s8(va0x1, va0x1, 4);
119 va1x0 = vext_s8(va1x0, va1x0, 4);
120 va1x1 = vext_s8(va1x1, va1x1, 4);
121 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
122 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
123 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
124 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
125 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
126 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
127 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
128 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
129 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
130 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
131 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
132 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
133 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
134 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
135 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
136 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
137 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
138 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
139 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
140 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
141 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
142 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
143 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
144 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
145 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
146 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
147 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
148 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
149
150 k -= 16 * sizeof(int8_t);
151 }
152 if (k != 0) {
153 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
154 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
155
156 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
157 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
158 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
160 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
162 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
163 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
164
165 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
166 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
167 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
168 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
169 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
170 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
171 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
172 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
173 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
174 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
175 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
176 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
177 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
178 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
179 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
180 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
181 va0x0 = vext_s8(va0x0, va0x0, 4);
182 va1x0 = vext_s8(va1x0, va1x0, 4);
183 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
184 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
185 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
186 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
187 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
188 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
189 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
190 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
191 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
192 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
193 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
194 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
195 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
196 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
197 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
198 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
199
200 }
201
202 p -= 2 * sizeof(void*);
203 } while (p != 0);
204
205 #if XNN_ARCH_ARM64
206 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
207 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
208 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
209 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
210 #else
211 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
212 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
213 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
214 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
215 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
216 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
217 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
218 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
219 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
220 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
221 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
222 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
223 #endif
224
225 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
226 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
227 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
228 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
229
230 const float32x4_t vscale0123 = vld1q_f32(w); w = (const float*) w + 4;
231 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
232 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
233 const float32x4_t vscale4567 = vld1q_f32(w); w = (const float*) w + 4;
234 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
235 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
236
237 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
238 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
239 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
240 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
241
242 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
243 #if XNN_ARCH_ARM64
244 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
245 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
246
247 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
248 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
249
250 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
251 #else
252 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
253 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
254
255 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
256 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
257
258 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
259 #endif
260
261 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
262 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
263
264 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max);
265 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
266
267 if (nc >= 8) {
268 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
269 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
270
271 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
272 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
273
274 a = (const int8_t**restrict) ((uintptr_t) a - ks);
275
276 nc -= 8;
277 } else {
278 if (nc & 4) {
279 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
280 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
281 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
282 }
283 if (nc & 2) {
284 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
285 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
286 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
287 }
288 if (nc & 1) {
289 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
290 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
291 }
292
293 nc = 0;
294 }
295 } while (nc != 0);
296 }
297