1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_fp32_ukernel_2x8c2s4__neon_mlal(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_fp32_ukernel_2x8c2s4__neon_mlal(
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 <= 2);
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 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 kc = round_up_po2(kc, 8 * sizeof(int8_t));
49 do {
50 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc1x0123 = vacc0x0123;
53 int32x4_t vacc1x4567 = vacc0x4567;
54
55 size_t k = kc;
56 while (k >= 16 * sizeof(int8_t)) {
57 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
58 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
59 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
60 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
61
62 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
63 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
64 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70
71 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
72 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
73 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
74 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
75 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
76 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
77 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
78 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
79 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
80 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
81 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
82 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
83 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
84 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
85 va0x0 = vext_s8(va0x0, va0x0, 2);
86 va0x1 = vext_s8(va0x1, va0x1, 2);
87 va1x0 = vext_s8(va1x0, va1x0, 2);
88 va1x1 = vext_s8(va1x1, va1x1, 2);
89 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
90 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
91 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
92 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
93 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
94 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
95 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
96 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
97 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
98 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
99 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
100 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
101 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
102 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
103 va0x0 = vext_s8(va0x0, va0x0, 2);
104 va0x1 = vext_s8(va0x1, va0x1, 2);
105 va1x0 = vext_s8(va1x0, va1x0, 2);
106 va1x1 = vext_s8(va1x1, va1x1, 2);
107 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
108 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
109 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
110 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
111 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
112 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
113 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
114 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
115 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
116 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
117 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
118 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
119 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
120 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
121 va0x0 = vext_s8(va0x0, va0x0, 2);
122 va0x1 = vext_s8(va0x1, va0x1, 2);
123 va1x0 = vext_s8(va1x0, va1x0, 2);
124 va1x1 = vext_s8(va1x1, va1x1, 2);
125 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
126 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
127 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
128 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
129 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
130 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
131 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
132 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
133 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
134 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
135 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
136 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
137 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
138 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
139
140 k -= 16 * sizeof(int8_t);
141 }
142 if (k != 0) {
143 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
144 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
145
146 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
147 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
148 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
149 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
150 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
151 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
152 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
153 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
154
155 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
156 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
157 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
158 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
159 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
160 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
161 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
162 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
163 va0x0 = vext_s8(va0x0, va0x0, 2);
164 va1x0 = vext_s8(va1x0, va1x0, 2);
165 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
166 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
167 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
168 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
169 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
170 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
171 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
172 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
173 va0x0 = vext_s8(va0x0, va0x0, 2);
174 va1x0 = vext_s8(va1x0, va1x0, 2);
175 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
176 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
177 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
178 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
179 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
180 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
181 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
182 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
183 va0x0 = vext_s8(va0x0, va0x0, 2);
184 va1x0 = vext_s8(va1x0, va1x0, 2);
185 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
186 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
187 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
188 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
189 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
190 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
191 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
192 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
193
194 }
195
196 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
197 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
198 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
199 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
200
201 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
202 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
203 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
204 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
205 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
206
207 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
208 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
209 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
210 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
211 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
212
213 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
214 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
215 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
216 vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
217 vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
218
219 #if XNN_ARCH_ARM64
220 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
221 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
222
223
224 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
225 #else
226 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
227 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
228
229
230 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
231 #endif
232
233 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
234 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
235
236 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
237 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
238
239 if (nc >= 8) {
240 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
241 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
242
243 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
244 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
245
246 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
247 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
248
249 nc -= 8;
250 } else {
251 // Final case where not all of the 8 columns fit in the destination.
252 if (nc & 4) {
253 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
254 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
255 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
256 }
257 if (nc & 2) {
258 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
259 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
260 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
261 }
262 if (nc & 1) {
263 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
264 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
265 }
266
267 nc = 0;
268 }
269 } while (nc != 0);
270 }
271