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/math.h>
16
xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r(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)])17 void xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const int8_t** restrict a,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const int8_t* zero,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (1 * sizeof(void*)) == 0);
37 assert(a_offset % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 2 * sizeof(int8_t));
43 int8_t* c0 = c;
44
45 do {
46 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
48
49 size_t p = ks;
50 do {
51 const int8_t* restrict a0 = a[0];
52 if XNN_UNPREDICTABLE(a0 != zero) {
53 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
54 }
55 a += 1;
56
57 size_t k = kc;
58
59 while (k >= 16 * sizeof(int8_t)) {
60 const int16x4x4_t va0x0 = vld4_dup_s16((const void*)a0); a0 += 8;
61 const int16x4x4_t va0x1 = vld4_dup_s16((const void*)a0); a0 += 8;
62
63 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71
72 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va0x0.val[0]);
73 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va0x1.val[0]);
74
75 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
76 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
78 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
79 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
80 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
82 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
83 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va0x0.val[1]);
84 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va0x1.val[1]);
85
86 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
87 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
89 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
90 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
91 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
93 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
94 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va0x0.val[2]);
95 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va0x1.val[2]);
96
97 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
98 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
100 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
101 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
102 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
104 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
105 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va0x0.val[3]);
106 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va0x1.val[3]);
107
108 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
109 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
111 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
112 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
113 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
115 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
116
117 k -= 16 * sizeof(int8_t);
118 }
119
120 if (k >= 8 * sizeof(int8_t)) {
121 const int16x4x4_t va0 = vld4_dup_s16((const void*)a0); a0 += 8;
122
123 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131
132 const int8x8_t va0c0 = vreinterpret_s8_s16(va0.val[0]);
133
134 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
135 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
136 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
137 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
138 const int8x8_t va0c1 = vreinterpret_s8_s16(va0.val[1]);
139
140 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
141 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
142 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
143 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
144 const int8x8_t va0c2 = vreinterpret_s8_s16(va0.val[2]);
145
146 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
147 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
148 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
149 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
150 const int8x8_t va0c3 = vreinterpret_s8_s16(va0.val[3]);
151
152 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
153 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
154 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
155 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
156
157 k -= 8 * sizeof(int8_t);
158 }
159
160 if XNN_UNLIKELY(k != 0) {
161 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
162
163 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165
166 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
167 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
168 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
169 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
170 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
171
172 if (k > 2 * sizeof(int8_t)) {
173 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175
176 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
177 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
178 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
179 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
180 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
181
182 if (k > 4 * sizeof(int8_t)) {
183 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185
186 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
187 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
188 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
189 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
190 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
191 }
192 }
193 }
194 p -= 1 * sizeof(void*);
195 } while (p != 0);
196
197 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
198 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
199
200 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
201 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
202 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
203
204 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
205 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
206 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
207
208 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
209 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
210 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
211
212 #if XNN_ARCH_ARM64
213 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
214
215
216 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
217 #else
218 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
219
220
221 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
222 #endif
223
224 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
225 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
226
227 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max);
228 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
229
230 if (nc >= 8) {
231 vst1_s8(c0 + 0, vout0x01234567);
232
233 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
234
235 a = (const int8_t**restrict) ((uintptr_t) a - ks);
236
237 nc -= 8;
238 } else {
239 if (nc & 4) {
240 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
241 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
242 }
243 if (nc & 2) {
244 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
245 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
246 }
247 if (nc & 1) {
248 vst1_lane_s8(c0, vout0x01234567, 0);
249 }
250
251 nc = 0;
252 }
253 } while (nc != 0);
254 }
255