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_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r(
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_qc8_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 int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
51 const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
52 const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
53 const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
54 const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
55 const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
56 const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
57 const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
58 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66
67 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
68 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
69
70 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
71 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
73 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
74 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
75 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
77 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
78 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
79 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
80
81 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
82 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
84 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
85 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
86 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
88 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
89 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
90 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
91
92 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
93 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
95 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
96 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
97 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
99 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
100 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
101 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
102
103 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
104 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
106 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
107 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
108 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
110 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
111
112 k -= 16 * sizeof(int8_t);
113 }
114
115 if (k >= 8 * sizeof(int8_t)) {
116 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
117 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
118 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
119 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
120
121 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129
130 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
131
132 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
133 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
134 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
135 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
136 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
137
138 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
139 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
140 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
141 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
142 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
143
144 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
145 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
146 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
147 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
148 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
149
150 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
151 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
152 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
153 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
154
155 k -= 8 * sizeof(int8_t);
156 }
157
158 if XNN_UNLIKELY(k != 0) {
159 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
160
161 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163
164 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
165 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
166 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
167 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
168 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
169
170 if (k > 2 * sizeof(int8_t)) {
171 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173
174 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
175 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
176 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
177 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
178 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
179
180 if (k > 4 * sizeof(int8_t)) {
181 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183
184 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
185 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
186 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
187 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
188 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
189 }
190 }
191 }
192
193 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
194 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
195
196 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
197 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
198 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
199 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
200
201 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
202 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
203
204 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
205 #if XNN_ARCH_ARM64
206 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
207
208 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
209
210 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
211 #else
212 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
213
214 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
215
216 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
217 #endif
218
219 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
220 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
221
222 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.output_max);
223 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
224
225 if (nc >= 8) {
226 vst1_s8(c0 + 0, vout0x01234567);
227
228 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
229
230 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
231
232 nc -= 8;
233 } else {
234 // Final case where not all of the 8 columns fit in the destination.
235 if (nc & 4) {
236 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
237 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
238 }
239 if (nc & 2) {
240 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
241 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
242 }
243 if (nc & 1) {
244 vst1_lane_s8(c0, vout0x01234567, 0);
245 }
246
247 nc = 0;
248 }
249 } while (nc != 0);
250 }
251