Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-5x16c4-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_qc4w_gemm_minmax_ukernel_5x16c4__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 struct xnn_f16_qc4w_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 <= 5);
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
68 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
69 // Loop over groups of 16 columns.
70 do {
71 // Initialize accumulators with bias. 16 bias values are loaded from the
72 // weight matrix, at the start of the group of 16 columns.
73 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
74 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
75 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
76 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
77 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
78 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
79 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
80 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
81 int32x4_t vacc0x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point01), 0);
82 int32x4_t vacc1x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point01), 0);
83 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
84 int32x4_t vacc0xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point01), 0);
85 int32x4_t vacc1xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point01), 0);
86 const int32x4_t vinput_zero_point23 = vld1q_s32(&quantization_params[2].zero_point);
87 int32x4_t vacc2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point23), 0);
88 int32x4_t vacc3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point23), 0);
89 int32x4_t vacc2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point23), 0);
90 int32x4_t vacc3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point23), 0);
91 int32x4_t vacc2x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point23), 0);
92 int32x4_t vacc3x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point23), 0);
93 int32x4_t vacc2xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point23), 0);
94 int32x4_t vacc3xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point23), 0);
95 const int32x4_t vinput_zero_point4 = vld1q_dup_s32(&quantization_params[4].zero_point);
96 int32x4_t vacc4x0123 = vmulq_s32(vksum0123, vinput_zero_point4);
97 int32x4_t vacc4x4567 = vmulq_s32(vksum4567, vinput_zero_point4);
98 int32x4_t vacc4x89AB = vmulq_s32(vksum89AB, vinput_zero_point4);
99 int32x4_t vacc4xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point4);
100
101 // Inner accumulation loop along the 16 columns.
102 size_t k = kc;
103 // 2x partial unrolled loop to load 8 bytes at a time.
104 while (k >= 8 * sizeof(int8_t)) {
105 // Load a 5x8 block of activations.
106 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
107 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
108 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
109 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
110 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
111
112 // Load a 8x16 block of weights.
113 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
114 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
115 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
116 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
117 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
118 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
119 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
120 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
121 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
122 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
123 const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
124 const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);
125
126 // Multiply-accumulate: 5x8 * 8x16 --> 5x16.
127 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
128 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
129 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
130 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
131 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
132 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
133 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
134 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
135 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
136 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
137 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
138 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
139 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
140 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
141 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
142 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
143 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
144 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
145 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
146 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
147 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
148 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
149 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
150 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
151 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
152 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
153 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
154 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
155 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
156 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
157 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
158 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
159 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
160 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
161 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
162 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
163 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
164 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
165 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
166 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
167
168 k -= 8 * sizeof(int8_t);
169 }
170 // Handle up to 4 final positions of `k`
171 if XNN_UNLIKELY(k != 0) {
172 // Load a 5x4 block of activations.
173 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
174 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
175 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
176 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
177 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
178
179 // Load a 4x16 block of weights.
180 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
181 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
182 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
183 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
184 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
185 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
186 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
187 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
188
189 // Multiply-accumulate: 5x4 * 4x16 --> 5x16.
190 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
191 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
192 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
193 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
194 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
195 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
196 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
197 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
198 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
199 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
200 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
201 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
202 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
203 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
204 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
205 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
206 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
207 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
208 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
209 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
210 }
211
212 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
213 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
214 float32x4_t vout0x89AB = vcvtq_n_f32_s32(vacc0x89AB, 4);
215 float32x4_t vout0xCDEF = vcvtq_n_f32_s32(vacc0xCDEF, 4);
216 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
217 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
218 float32x4_t vout1x89AB = vcvtq_n_f32_s32(vacc1x89AB, 4);
219 float32x4_t vout1xCDEF = vcvtq_n_f32_s32(vacc1xCDEF, 4);
220 float32x4_t vout2x0123 = vcvtq_n_f32_s32(vacc2x0123, 4);
221 float32x4_t vout2x4567 = vcvtq_n_f32_s32(vacc2x4567, 4);
222 float32x4_t vout2x89AB = vcvtq_n_f32_s32(vacc2x89AB, 4);
223 float32x4_t vout2xCDEF = vcvtq_n_f32_s32(vacc2xCDEF, 4);
224 float32x4_t vout3x0123 = vcvtq_n_f32_s32(vacc3x0123, 4);
225 float32x4_t vout3x4567 = vcvtq_n_f32_s32(vacc3x4567, 4);
226 float32x4_t vout3x89AB = vcvtq_n_f32_s32(vacc3x89AB, 4);
227 float32x4_t vout3xCDEF = vcvtq_n_f32_s32(vacc3xCDEF, 4);
228 float32x4_t vout4x0123 = vcvtq_n_f32_s32(vacc4x0123, 4);
229 float32x4_t vout4x4567 = vcvtq_n_f32_s32(vacc4x4567, 4);
230 float32x4_t vout4x89AB = vcvtq_n_f32_s32(vacc4x89AB, 4);
231 float32x4_t vout4xCDEF = vcvtq_n_f32_s32(vacc4xCDEF, 4);
232 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
233 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
234 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
235 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
236 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
237 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
238 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
239 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
240 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
241 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
242 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
243 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
244 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
245 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
246 vout2x89AB = vmulq_lane_f32(vout2x89AB, vget_low_f32(vinput_scale23), 1);
247 vout3x89AB = vmulq_lane_f32(vout3x89AB, vget_high_f32(vinput_scale23), 1);
248 vout2xCDEF = vmulq_lane_f32(vout2xCDEF, vget_low_f32(vinput_scale23), 1);
249 vout3xCDEF = vmulq_lane_f32(vout3xCDEF, vget_high_f32(vinput_scale23), 1);
250 const float32x4_t vinput_scale4 = vld1q_dup_f32(&quantization_params[4].inv_scale);
251 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale4);
252 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale4);
253 vout4x89AB = vmulq_f32(vout4x89AB, vinput_scale4);
254 vout4xCDEF = vmulq_f32(vout4xCDEF, vinput_scale4);
255
256 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
257 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
258 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
259 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
260
261 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
262 #if XNN_ARCH_ARM64
263 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
264 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
265 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
266 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
267 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
268 #else
269 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
270 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
271 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
272 vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
273 vout4x0123 = vmlaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
274 #endif
275 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
276 #if XNN_ARCH_ARM64
277 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
278 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
279 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
280 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
281 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
282 #else
283 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
284 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
285 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
286 vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
287 vout4x4567 = vmlaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
288 #endif
289 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
290 #if XNN_ARCH_ARM64
291 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
292 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
293 vout2x89AB = vfmaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
294 vout3x89AB = vfmaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
295 vout4x89AB = vfmaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
296 #else
297 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
298 vout1x89AB = vmlaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
299 vout2x89AB = vmlaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
300 vout3x89AB = vmlaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
301 vout4x89AB = vmlaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
302 #endif
303 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
304 #if XNN_ARCH_ARM64
305 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
306 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
307 vout2xCDEF = vfmaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
308 vout3xCDEF = vfmaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
309 vout4xCDEF = vfmaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
310 #else
311 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
312 vout1xCDEF = vmlaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
313 vout2xCDEF = vmlaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
314 vout3xCDEF = vmlaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
315 vout4xCDEF = vmlaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
316 #endif
317
318 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
319 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
320 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
321 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
322 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
323 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
324 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
325 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
326 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
327 float16x8_t vfp16out4x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout4x89AB), vcvt_f16_f32(vout4xCDEF));
328
329 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
330 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
331 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
332 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
333 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
334 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
335 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
336 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
337 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
338 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
339 vfp16out4x89ABCDEF = vmaxq_f16(vfp16out4x89ABCDEF, voutput_min);
340 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
341 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
342 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
343 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
344 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
345 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
346 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
347 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
348 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
349 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
350 vfp16out4x89ABCDEF = vminq_f16(vfp16out4x89ABCDEF, voutput_max);
351 if XNN_LIKELY(nc >= 16) {
352 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
353 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
354 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
355 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
356 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
357 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
358 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
359 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
360 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
361 vst1q_u16(c4 + 8, vreinterpretq_u16_f16(vfp16out4x89ABCDEF));
362
363 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
364 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
365 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
366 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
367 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
368
369 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
370 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
371 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
372 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
373 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
374
375 nc -= 16;
376 } else {
377 if (nc & 8) {
378 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
379 vfp16out0x01234567 = vfp16out0x89ABCDEF;
380 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
381 vfp16out1x01234567 = vfp16out1x89ABCDEF;
382 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
383 vfp16out2x01234567 = vfp16out2x89ABCDEF;
384 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
385 vfp16out3x01234567 = vfp16out3x89ABCDEF;
386 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567)); c4 += 8;
387 vfp16out4x01234567 = vfp16out4x89ABCDEF;
388 }
389 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
390 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
391 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
392 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
393 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
394 if (nc & 4) {
395 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
396 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
397 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
398 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
399 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
400 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
401 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
402 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
403 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
404 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
405 }
406 if (nc & 2) {
407 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
408 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
409 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
410 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
411 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
412 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
413 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
414 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
415 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
416 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
417 }
418 if (nc & 1) {
419 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
420 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
421 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
422 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
423 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
424 }
425 nc = 0;
426 }
427 } while (nc != 0);
428 }