1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-neon-mull-dup.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
xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r(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)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r(
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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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 kc = round_up_po2(kc, 2 * sizeof(int8_t));
44 int8_t* c0 = c;
45
46 do {
47 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
48 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
49
50 size_t p = ks;
51 do {
52 const int8_t* restrict a0 = a[0];
53 if XNN_UNPREDICTABLE(a0 != zero) {
54 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
55 }
56 a += 1;
57
58 size_t k = kc;
59
60 while (k >= 16 * sizeof(int8_t)) {
61 const int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
62 const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
63 const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
64 const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
65 const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
66 const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
67 const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
68 const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
69
70 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78
79 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
80 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
81
82 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
83 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
85 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
86 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
87 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
89 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
90 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
91 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
92
93 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
94 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
96 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
97 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
98 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
100 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
101 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
102 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
103
104 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
105 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
107 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
108 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
109 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
111 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
112 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
113 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
114
115 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
116 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
118 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
119 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
120 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
122 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
123
124 k -= 16 * sizeof(int8_t);
125 }
126
127 if (k >= 8 * sizeof(int8_t)) {
128 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
129 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
130 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
131 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
132
133 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141
142 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
143
144 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
145 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
146 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
147 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
148 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
149
150 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
151 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
152 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
153 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
154 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
155
156 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
157 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
158 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
159 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
160 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
161
162 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
163 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
164 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
165 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
166
167 k -= 8 * sizeof(int8_t);
168 }
169
170 if XNN_UNLIKELY(k != 0) {
171 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
172
173 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175
176 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
177 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
178 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
179 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
180 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
181
182 if (k > 2 * sizeof(int8_t)) {
183 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185
186 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
187 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
188 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
189 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
190 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
191
192 if (k > 4 * sizeof(int8_t)) {
193 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
194 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
195
196 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
197 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
198 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
199 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
200 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
201 }
202 }
203 }
204 p -= 1 * sizeof(void*);
205 } while (p != 0);
206
207 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
208 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
209
210 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
211 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
212 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
213 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
214
215 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
216 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
217
218 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
219 #if XNN_ARCH_ARM64
220 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
221
222 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
223
224 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
225 #else
226 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
227
228 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
229
230 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
231 #endif
232
233 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
234 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
235
236 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.output_max);
237 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
238
239 if (nc >= 8) {
240 vst1_s8(c0 + 0, vout0x01234567);
241
242 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
243
244 a = (const int8_t**restrict) ((uintptr_t) a - ks);
245
246 nc -= 8;
247 } else {
248 if (nc & 4) {
249 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
250 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
251 }
252 if (nc & 2) {
253 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
254 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
255 }
256 if (nc & 1) {
257 vst1_lane_s8(c0, vout0x01234567, 0);
258 }
259
260 nc = 0;
261 }
262 } while (nc != 0);
263 }
264