Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-5x8c8-minmax-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-neoni8mm.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_f32_qc8w_gemm_minmax_ukernel_5x8c8__neoni8mm(
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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_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, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((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 float* c2 = (float*) ((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 float* c3 = (float*) ((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 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67
68 // Loop over groups of 8 columns.
69 do {
70 // Initialize accumulators with bias. 8 bias values are loaded from the
71 // weight matrix, at the start of the group of 8 columns.
72 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
73 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
74 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
75 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
76 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
77 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
78 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
79 const int32x4_t vinput_zero_point23 = vld1q_s32(&quantization_params[2].zero_point);
80 const int32x4_t vksumzp2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point23), 0);
81 const int32x4_t vksumzp3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point23), 0);
82 const int32x4_t vksumzp2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point23), 0);
83 const int32x4_t vksumzp3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point23), 0);
84 const int32x4_t vinput_zero_point45 = vld1q_dup_s32(&quantization_params[4].zero_point);
85 const int32x4_t vksumzp4x0123 = vmulq_s32(vksum0123, vinput_zero_point45);
86 const int32x4_t vksumzp4x4567 = vmulq_s32(vksum4567, vinput_zero_point45);
87
88 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
89 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
90 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
91 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
92 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
93 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
94 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
95 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
96 int32x4_t vacc45x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp4x0123)));
97 int32x4_t vacc45x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp4x0123)));
98 int32x4_t vacc45x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp4x4567)));
99 int32x4_t vacc45x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp4x4567)));
100
101 // Inner accumulation loop along the 8 columns.
102 size_t k = kc;
103 // 2x partial unrolled loop to load 8 bytes at a time.
104
105 uint64x2x2_t va01x0123456789ABCDEF;
106 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
107 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
108 uint64x2x2_t va23x0123456789ABCDEF;
109 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
110 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
111 uint64x2x2_t va45x0123456789ABCDEF;
112 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
113 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
114
115 while (k >= 16 * sizeof(int8_t)) {
116 // Load a 5x16 block of activations.
117 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
118 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
119 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
120 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
121 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
122
123 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
124 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
125 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
126 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
127 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
128 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
129
130 // Load a 16x8 block of weights.
131 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
132 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
133 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
134 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
135 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
136 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
137 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
138 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
139
140 // Multiply-accumulate: 5x8 * 8x8 --> 5x8.
141 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
142 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
143 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
144 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
145 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
146 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
147 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
148 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
149 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
150 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
151 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
152 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
153 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
154 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
155 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
156 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
157 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
158 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
159 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
160 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
161 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
162 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
163 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
164 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
165
166 k -= 16 * sizeof(int8_t);
167 }
168 // Handle up to 8 final positions of `k`
169 if XNN_UNLIKELY(k != 0) {
170 // Load a 5x8 block of activations.
171 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
172 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
173 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
174 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
175 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
176
177 // Load a 16x8 block of weights.
178 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
179 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
180 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
181 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
182
183 // Multiply-accumulate: 5x4 * 4x8 --> 5x8.
184 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
185 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
186 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
187 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
188 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
189 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
190 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
191 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
192 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
193 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
194 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
195 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
196 }
197
198 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
199 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
200 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
201 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
202 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
203 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
204 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
205 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
206 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
207 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
208 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
209 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
210 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
211 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
212 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
213 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
214 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
215 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
216 float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123);
217 float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567);
218
219 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
220 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
221 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
222 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
223 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
224 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
225 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
226 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
227 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
228 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
229 const float32x4_t vinput_scale4 = vld1q_dup_f32(&quantization_params[4].inv_scale);
230 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale4);
231 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale4);
232
233 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
234 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
235
236 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
237 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
238 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
239 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
240 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
241 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
242 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
243 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
244 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
245 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
246 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
247 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
248
249 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
250 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
251 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
252 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
253 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
254 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
255 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
256 vout3x0123 = vmaxq_f32(vout3x0123, voutput_min);
257 vout3x4567 = vmaxq_f32(vout3x4567, voutput_min);
258 vout4x0123 = vmaxq_f32(vout4x0123, voutput_min);
259 vout4x4567 = vmaxq_f32(vout4x4567, voutput_min);
260
261 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
262 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
263 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
264 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
265 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
266 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
267 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
268 vout3x0123 = vminq_f32(vout3x0123, voutput_max);
269 vout3x4567 = vminq_f32(vout3x4567, voutput_max);
270 vout4x0123 = vminq_f32(vout4x0123, voutput_max);
271 vout4x4567 = vminq_f32(vout4x4567, voutput_max);
272
273 if XNN_LIKELY(nc >= 8) {
274 vst1q_f32(c0, vout0x0123);
275 vst1q_f32(c0 + 4, vout0x4567);
276 vst1q_f32(c1, vout1x0123);
277 vst1q_f32(c1 + 4, vout1x4567);
278 vst1q_f32(c2, vout2x0123);
279 vst1q_f32(c2 + 4, vout2x4567);
280 vst1q_f32(c3, vout3x0123);
281 vst1q_f32(c3 + 4, vout3x4567);
282 vst1q_f32(c4, vout4x0123);
283 vst1q_f32(c4 + 4, vout4x4567);
284
285 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
286 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
287 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
288 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
289 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
290
291 c0 = (float*) ((uintptr_t) c0 + cn_stride);
292 c1 = (float*) ((uintptr_t) c1 + cn_stride);
293 c2 = (float*) ((uintptr_t) c2 + cn_stride);
294 c3 = (float*) ((uintptr_t) c3 + cn_stride);
295 c4 = (float*) ((uintptr_t) c4 + cn_stride);
296
297 nc -= 8;
298 } else {
299 if (nc & 4) {
300 vst1q_f32(c0, vout0x0123); c0 += 4;
301 vout0x0123 = vout0x4567;
302 vst1q_f32(c1, vout1x0123); c1 += 4;
303 vout1x0123 = vout1x4567;
304 vst1q_f32(c2, vout2x0123); c2 += 4;
305 vout2x0123 = vout2x4567;
306 vst1q_f32(c3, vout3x0123); c3 += 4;
307 vout3x0123 = vout3x4567;
308 vst1q_f32(c4, vout4x0123); c4 += 4;
309 vout4x0123 = vout4x4567;
310 }
311 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
312 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
313 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
314 float32x2_t vout3x01 = vget_low_f32(vout3x0123);
315 float32x2_t vout4x01 = vget_low_f32(vout4x0123);
316 if (nc & 2) {
317 vst1_f32(c0, vout0x01); c0 += 2;
318 vst1_f32(c1, vout1x01); c1 += 2;
319 vst1_f32(c2, vout2x01); c2 += 2;
320 vst1_f32(c3, vout3x01); c3 += 2;
321 vst1_f32(c4, vout4x01); c4 += 2;
322 vout0x01 = vget_high_f32(vout0x0123);
323 vout1x01 = vget_high_f32(vout1x0123);
324 vout2x01 = vget_high_f32(vout2x0123);
325 vout3x01 = vget_high_f32(vout3x0123);
326 vout4x01 = vget_high_f32(vout4x0123);
327 }
328 if (nc & 1) {
329 vst1_lane_f32(c0, vout0x01, 0);
330 vst1_lane_f32(c1, vout1x01, 0);
331 vst1_lane_f32(c2, vout2x01, 0);
332 vst1_lane_f32(c3, vout3x01, 0);
333 vst1_lane_f32(c4, vout4x01, 0);
334 }
335 nc = 0;
336 }
337 } while (nc != 0);
338 }