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/math.h>
16
17
xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4s2__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_2x8c4s2__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 vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
52 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
53 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
54 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
55 int32x4_t vacc1x01 = vacc0x01;
56 int32x4_t vacc1x23 = vacc0x23;
57 int32x4_t vacc1x45 = vacc0x45;
58 int32x4_t vacc1x67 = vacc0x67;
59
60 size_t p = ks;
61 do {
62 const int8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const int8_t* restrict a1 = a[1];
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 a += 2;
71
72 size_t k = kc;
73 while (k >= 16 * sizeof(int8_t)) {
74 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
75 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
76 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
77 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
78
79 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87
88 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
89 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
90 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
91 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
92 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
93 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
94 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
95 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
96 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
97 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
98 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
99 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
100 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
101 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
102 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
103 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
104 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
105 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
106 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
107 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
108 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
109 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
110 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
111 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
112 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
113 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
114 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
115 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
116 va0x0 = vext_s8(va0x0, va0x0, 4);
117 va0x1 = vext_s8(va0x1, va0x1, 4);
118 va1x0 = vext_s8(va1x0, va1x0, 4);
119 va1x1 = vext_s8(va1x1, va1x1, 4);
120 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
121 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
122 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
123 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
124 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
125 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
126 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
127 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
128 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
129 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
130 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
131 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
132 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
133 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
134 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
135 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
136 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
137 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
138 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
139 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
140 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
141 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
142 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
143 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
144 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
145 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
146 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
147 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
148
149 k -= 16 * sizeof(int8_t);
150 }
151 if (k != 0) {
152 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
153 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
154
155 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
156 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
157 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
158 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
160 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
162 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
163
164 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
165 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
166 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
167 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
168 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
169 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
170 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
171 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
172 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
173 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
174 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
175 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
176 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
177 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
178 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
179 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
180 va0x0 = vext_s8(va0x0, va0x0, 4);
181 va1x0 = vext_s8(va1x0, va1x0, 4);
182 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
183 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
184 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
185 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
186 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
187 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
188 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
189 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
190 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
191 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
192 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
193 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
194 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
195 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
196 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
197 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
198
199 }
200
201 p -= 2 * sizeof(void*);
202 } while (p != 0);
203
204 #if XNN_ARCH_ARM64
205 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
206 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
207 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
208 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
209 #else
210 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
211 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
212 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
213 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
214 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
215 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
216 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
217 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
218 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
219 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
220 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
221 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
222 #endif
223
224 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
225 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
226 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
227 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
228
229 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
230 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
231 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
232 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
233 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
234
235 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
236 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
237 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
238 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
239 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
240
241 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
242 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
243 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
244 vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
245 vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
246
247 #if XNN_ARCH_ARM64
248 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
249 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
250
251
252 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
253 #else
254 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
255 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
256
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_neon.output_min);
262 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
263
264 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.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