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/math.h>
16
17
xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neon_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)])18 void xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 int8_t* c0 = c;
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 c1 = c0;
47 }
48
49 kc = round_up_po2(kc, 8 * sizeof(int8_t));
50 do {
51 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc1x0123 = vacc0x0123;
54 int32x4_t vacc1x4567 = vacc0x4567;
55
56 size_t p = ks;
57 do {
58 const int8_t* restrict a0 = a[0];
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 const int8_t* restrict a1 = a[1];
63 if XNN_UNPREDICTABLE(a1 != zero) {
64 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
65 }
66 a += 2;
67
68 size_t k = kc;
69 while (k >= 16 * sizeof(int8_t)) {
70 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
71 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
72 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
73 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
74
75 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83
84 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
85 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
86 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
87 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
88 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
89 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
90 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
91 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
92 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
93 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
94 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
95 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
96 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
97 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
98 va0x0 = vext_s8(va0x0, va0x0, 2);
99 va0x1 = vext_s8(va0x1, va0x1, 2);
100 va1x0 = vext_s8(va1x0, va1x0, 2);
101 va1x1 = vext_s8(va1x1, va1x1, 2);
102 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
103 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
104 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
105 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
106 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
107 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
108 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
109 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
110 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
111 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
112 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
113 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
114 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
115 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
116 va0x0 = vext_s8(va0x0, va0x0, 2);
117 va0x1 = vext_s8(va0x1, va0x1, 2);
118 va1x0 = vext_s8(va1x0, va1x0, 2);
119 va1x1 = vext_s8(va1x1, va1x1, 2);
120 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
121 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
122 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
123 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
124 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
125 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
126 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
127 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
128 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
129 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
130 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
131 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
132 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
133 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
134 va0x0 = vext_s8(va0x0, va0x0, 2);
135 va0x1 = vext_s8(va0x1, va0x1, 2);
136 va1x0 = vext_s8(va1x0, va1x0, 2);
137 va1x1 = vext_s8(va1x1, va1x1, 2);
138 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
139 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
140 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
141 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
142 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
143 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
144 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
145 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
146 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
147 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
148 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
149 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
150 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
151 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
152
153 k -= 16 * sizeof(int8_t);
154 }
155 if (k != 0) {
156 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
157 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
158
159 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
160 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
162 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
163 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
164 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
165 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
166 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
167
168 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
169 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
170 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
171 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
172 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
173 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
174 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
175 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
176 va0x0 = vext_s8(va0x0, va0x0, 2);
177 va1x0 = vext_s8(va1x0, va1x0, 2);
178 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
179 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
180 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
181 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
182 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
183 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
184 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
185 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
186 va0x0 = vext_s8(va0x0, va0x0, 2);
187 va1x0 = vext_s8(va1x0, va1x0, 2);
188 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
189 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
190 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
191 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
192 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
193 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
194 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
195 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
196 va0x0 = vext_s8(va0x0, va0x0, 2);
197 va1x0 = vext_s8(va1x0, va1x0, 2);
198 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
199 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
200 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
201 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
202 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
203 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
204 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
205 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
206
207 }
208
209 p -= 2 * sizeof(void*);
210 } while (p != 0);
211
212 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
213 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
214 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
215 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
216
217 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
218 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
219 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
220 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
221 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
222
223 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
224 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
225 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
226 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
227 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
228
229 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
230 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
231 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
232 vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
233 vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
234
235 #if XNN_ARCH_ARM64
236 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
237 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
238
239
240 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
241 #else
242 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
243 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
244
245
246 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
247 #endif
248
249 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
250 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
251
252 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
253 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
254
255 if (nc >= 8) {
256 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
257 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
258
259 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
260 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
261
262 a = (const int8_t**restrict) ((uintptr_t) a - ks);
263
264 nc -= 8;
265 } else {
266 if (nc & 4) {
267 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
268 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
269 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
270 }
271 if (nc & 2) {
272 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
273 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
274 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
275 }
276 if (nc & 1) {
277 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
278 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
279 }
280
281 nc = 0;
282 }
283 } while (nc != 0);
284 }
285