Debian Codegraph @

src/qd8-f16-qc8w-gemm/gen/qd8-f16-qc8w-gemm-6x16c4-minmax-neondotfp16arith.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-neondot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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
18 void xnn_qd8_f16_qc8w_gemm_minmax_ukernel_6x16c4__neondotfp16arith(
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 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 6);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 4 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 uint16_t* c0 = (uint16_t*) c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr != 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73
74 // Loop over groups of 16 columns.
75 do {
76 // Initialize accumulators with bias. 16 bias values are loaded from the
77 // weight matrix, at the start of the group of 16 columns.
78 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
79 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
80 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
81 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
82 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
83 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
84 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
85 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
86 int32x4_t vacc0x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point01), 0);
87 int32x4_t vacc1x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point01), 0);
88 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
89 int32x4_t vacc0xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point01), 0);
90 int32x4_t vacc1xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point01), 0);
91 const int32x4_t vinput_zero_point23 = vld1q_s32(&quantization_params[2].zero_point);
92 int32x4_t vacc2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point23), 0);
93 int32x4_t vacc3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point23), 0);
94 int32x4_t vacc2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point23), 0);
95 int32x4_t vacc3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point23), 0);
96 int32x4_t vacc2x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point23), 0);
97 int32x4_t vacc3x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point23), 0);
98 int32x4_t vacc2xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point23), 0);
99 int32x4_t vacc3xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point23), 0);
100 const int32x4_t vinput_zero_point45 = vld1q_s32(&quantization_params[4].zero_point);
101 int32x4_t vacc4x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point45), 0);
102 int32x4_t vacc5x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point45), 0);
103 int32x4_t vacc4x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point45), 0);
104 int32x4_t vacc5x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point45), 0);
105 int32x4_t vacc4x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point45), 0);
106 int32x4_t vacc5x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point45), 0);
107 int32x4_t vacc4xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point45), 0);
108 int32x4_t vacc5xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point45), 0);
109
110 // Inner accumulation loop along the 16 columns.
111 size_t k = kc;
112 // 2x partial unrolled loop to load 8 bytes at a time.
113 while (k >= 8 * sizeof(int8_t)) {
114 // Load a 6x8 block of activations.
115 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
116 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
117 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
118 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
119 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
120 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
121
122 // Load a 8x16 block of weights.
123 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
124 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
125 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
126 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
127 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
128 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
129 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
130 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
131
132 // Multiply-accumulate: 6x8 * 8x16 --> 6x16.
133 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
134 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
135 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
136 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
137 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
138 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
139 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
140 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
141 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
142 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
143 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
144 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
145 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
146 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
147 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
148 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
149 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
150 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
151 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
152 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
153 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
154 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
155 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
156 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
157 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
158 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
159 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
160 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
161 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
162 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
163 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
164 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
165 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
166 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
167 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
168 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
169 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
170 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
171 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
172 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
173 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
174 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
175 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
176 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
177 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
178 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
179 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1);
180 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1);
181
182 k -= 8 * sizeof(int8_t);
183 }
184 // Handle up to 4 final positions of `k`
185 if XNN_UNLIKELY(k != 0) {
186 // Load a 6x4 block of activations.
187 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
188 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
189 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
190 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
191 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
192 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4;
193
194 // Load a 4x16 block of weights.
195 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
196 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
197 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
198 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
199
200 // Multiply-accumulate: 6x4 * 4x16 --> 6x16.
201 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
202 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
203 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
204 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
205 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
206 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
207 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
208 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
209 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
210 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
211 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
212 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
213 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
214 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
215 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
216 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
217 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
218 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
219 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
220 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
221 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
222 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
223 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
224 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
225 }
226
227 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
228 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
229 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
230 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
231 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
232 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
233 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
234 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
235 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
236 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
237 float32x4_t vout2x89AB = vcvtq_f32_s32(vacc2x89AB);
238 float32x4_t vout2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
239 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
240 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
241 float32x4_t vout3x89AB = vcvtq_f32_s32(vacc3x89AB);
242 float32x4_t vout3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
243 float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123);
244 float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567);
245 float32x4_t vout4x89AB = vcvtq_f32_s32(vacc4x89AB);
246 float32x4_t vout4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
247 float32x4_t vout5x0123 = vcvtq_f32_s32(vacc5x0123);
248 float32x4_t vout5x4567 = vcvtq_f32_s32(vacc5x4567);
249 float32x4_t vout5x89AB = vcvtq_f32_s32(vacc5x89AB);
250 float32x4_t vout5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
251 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
252 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
253 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
254 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
255 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
256 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
257 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
258 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
259 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
260 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
261 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
262 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
263 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
264 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
265 vout2x89AB = vmulq_lane_f32(vout2x89AB, vget_low_f32(vinput_scale23), 1);
266 vout3x89AB = vmulq_lane_f32(vout3x89AB, vget_high_f32(vinput_scale23), 1);
267 vout2xCDEF = vmulq_lane_f32(vout2xCDEF, vget_low_f32(vinput_scale23), 1);
268 vout3xCDEF = vmulq_lane_f32(vout3xCDEF, vget_high_f32(vinput_scale23), 1);
269 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
270 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
271 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
272 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
273 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
274 vout4x89AB = vmulq_lane_f32(vout4x89AB, vget_low_f32(vinput_scale45), 1);
275 vout5x89AB = vmulq_lane_f32(vout5x89AB, vget_high_f32(vinput_scale45), 1);
276 vout4xCDEF = vmulq_lane_f32(vout4xCDEF, vget_low_f32(vinput_scale45), 1);
277 vout5xCDEF = vmulq_lane_f32(vout5xCDEF, vget_high_f32(vinput_scale45), 1);
278
279 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
280 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
281 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
282 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
283
284 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
285 #if XNN_ARCH_ARM64
286 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
287 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
288 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
289 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
290 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
291 vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
292 #else
293 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
294 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
295 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
296 vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
297 vout4x0123 = vmlaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
298 vout5x0123 = vmlaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
299 #endif
300 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
301 #if XNN_ARCH_ARM64
302 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
303 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
304 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
305 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
306 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
307 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
308 #else
309 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
310 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
311 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
312 vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
313 vout4x4567 = vmlaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
314 vout5x4567 = vmlaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
315 #endif
316 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
317 #if XNN_ARCH_ARM64
318 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
319 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
320 vout2x89AB = vfmaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
321 vout3x89AB = vfmaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
322 vout4x89AB = vfmaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
323 vout5x89AB = vfmaq_f32(vbias89AB, vout5x89AB, vfilter_output_scale89AB);
324 #else
325 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
326 vout1x89AB = vmlaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
327 vout2x89AB = vmlaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
328 vout3x89AB = vmlaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
329 vout4x89AB = vmlaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
330 vout5x89AB = vmlaq_f32(vbias89AB, vout5x89AB, vfilter_output_scale89AB);
331 #endif
332 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
333 #if XNN_ARCH_ARM64
334 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
335 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
336 vout2xCDEF = vfmaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
337 vout3xCDEF = vfmaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
338 vout4xCDEF = vfmaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
339 vout5xCDEF = vfmaq_f32(vbiasCDEF, vout5xCDEF, vfilter_output_scaleCDEF);
340 #else
341 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
342 vout1xCDEF = vmlaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
343 vout2xCDEF = vmlaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
344 vout3xCDEF = vmlaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
345 vout4xCDEF = vmlaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
346 vout5xCDEF = vmlaq_f32(vbiasCDEF, vout5xCDEF, vfilter_output_scaleCDEF);
347 #endif
348
349 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
350 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
351 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
352 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
353 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
354 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
355 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
356 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
357 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
358 float16x8_t vfp16out4x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout4x89AB), vcvt_f16_f32(vout4xCDEF));
359 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
360 float16x8_t vfp16out5x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout5x89AB), vcvt_f16_f32(vout5xCDEF));
361
362 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
363 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
364 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
365 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
366 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
367 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
368 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
369 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
370 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
371 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
372 vfp16out4x89ABCDEF = vmaxq_f16(vfp16out4x89ABCDEF, voutput_min);
373 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
374 vfp16out5x89ABCDEF = vmaxq_f16(vfp16out5x89ABCDEF, voutput_min);
375 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
376 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
377 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
378 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
379 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
380 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
381 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
382 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
383 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
384 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
385 vfp16out4x89ABCDEF = vminq_f16(vfp16out4x89ABCDEF, voutput_max);
386 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
387 vfp16out5x89ABCDEF = vminq_f16(vfp16out5x89ABCDEF, voutput_max);
388 if XNN_LIKELY(nc >= 16) {
389 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
390 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
391 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
392 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
393 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
394 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
395 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
396 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
397 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
398 vst1q_u16(c4 + 8, vreinterpretq_u16_f16(vfp16out4x89ABCDEF));
399 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
400 vst1q_u16(c5 + 8, vreinterpretq_u16_f16(vfp16out5x89ABCDEF));
401
402 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
403 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
404 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
405 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
406 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
407 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
408
409 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
410 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
411 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
412 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
413 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
414 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
415
416 nc -= 16;
417 } else {
418 if (nc & 8) {
419 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
420 vfp16out0x01234567 = vfp16out0x89ABCDEF;
421 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
422 vfp16out1x01234567 = vfp16out1x89ABCDEF;
423 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
424 vfp16out2x01234567 = vfp16out2x89ABCDEF;
425 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
426 vfp16out3x01234567 = vfp16out3x89ABCDEF;
427 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567)); c4 += 8;
428 vfp16out4x01234567 = vfp16out4x89ABCDEF;
429 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567)); c5 += 8;
430 vfp16out5x01234567 = vfp16out5x89ABCDEF;
431 }
432 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
433 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
434 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
435 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
436 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
437 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
438 if (nc & 4) {
439 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
440 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
441 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
442 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
443 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
444 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
445 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
446 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
447 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
448 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
449 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
450 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
451 }
452 if (nc & 2) {
453 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
454 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
455 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
456 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
457 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
458 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
459 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
460 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
461 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
462 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
463 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
464 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
465 }
466 if (nc & 1) {
467 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
468 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
469 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
470 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
471 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
472 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
473 }
474 nc = 0;
475 }
476 } while (nc != 0);
477 }