1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup(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_1x8c2__neonv8_mlal_dup(
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 <= 1);
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 kc = round_up_po2(kc, 2 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42
43 do {
44 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
46
47 size_t k = kc;
48
49 while (k >= 16 * sizeof(int8_t)) {
50 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
51 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
52 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
53 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
54 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
55 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60
61 const int8x8_t va0c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0));
62 const int8x8_t va0c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0));
63
64 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
65 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
67 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
68 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
69 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
71 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
72 const int8x8_t va0c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1));
73 const int8x8_t va0c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1));
74
75 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
76 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
78 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
79 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
80 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
82 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
83 const int8x8_t va0c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2));
84 const int8x8_t va0c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2));
85
86 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
87 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
89 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
90 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
91 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
93 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
94 const int8x8_t va0c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3));
95 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3));
96
97 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
98 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
100 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
101 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
102 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
104 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
105
106 k -= 16 * sizeof(int8_t);
107 }
108
109 if (k >= 8 * sizeof(int8_t)) {
110 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
111
112 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120
121 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
122
123 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
124 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
125 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
126 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
127 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
128
129 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
130 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
131 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
132 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
133 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
134
135 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
136 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
137 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
138 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
139 const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
140
141 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
142 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
143 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
144 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
145
146 k -= 8 * sizeof(int8_t);
147 }
148
149 if XNN_UNLIKELY(k != 0) {
150 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
151
152 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154
155 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
156 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
157 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
158 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
159 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
160
161 if (k > 2 * sizeof(int8_t)) {
162 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164
165 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
166 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
167 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
168 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
169 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
170
171 if (k > 4 * sizeof(int8_t)) {
172 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174
175 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
176 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
177 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
178 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
179 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
180 }
181 }
182 }
183
184 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
185 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
186
187 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
188 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
189 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
190
191 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
192 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
193
194 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
195 #if XNN_ARCH_ARM64
196 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
197
198 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
199
200 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
201 #else
202 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
203
204 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
205
206 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
207 #endif
208
209 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
210 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
211
212 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.output_max);
213 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
214
215 if (nc >= 8) {
216 vst1_s8(c0 + 0, vout0x01234567);
217
218 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
219
220 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
221
222 nc -= 8;
223 } else {
224 // Final case where not all of the 8 columns fit in the destination.
225 if (nc & 4) {
226 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
227 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
228 }
229 if (nc & 2) {
230 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
231 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
232 }
233 if (nc & 1) {
234 vst1_lane_s8(c0, vout0x01234567, 0);
235 }
236
237 nc = 0;
238 }
239 } while (nc != 0);
240 }
241