Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-7x8c8-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_7x8c8__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 <= 7);
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 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr < 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
74 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
75 if XNN_UNPREDICTABLE(mr <= 6) {
76 a6 = a5;
77 c6 = c5;
78 }
79
80 // Loop over groups of 8 columns.
81 do {
82 // Initialize accumulators with bias. 8 bias values are loaded from the
83 // weight matrix, at the start of the group of 8 columns.
84 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
85 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
86 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
87 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
88 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
89 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
90 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
91 const int32x4_t vinput_zero_point23 = vld1q_s32(&quantization_params[2].zero_point);
92 const int32x4_t vksumzp2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point23), 0);
93 const int32x4_t vksumzp3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point23), 0);
94 const int32x4_t vksumzp2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point23), 0);
95 const int32x4_t vksumzp3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point23), 0);
96 const int32x4_t vinput_zero_point45 = vld1q_s32(&quantization_params[4].zero_point);
97 const int32x4_t vksumzp4x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point45), 0);
98 const int32x4_t vksumzp5x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point45), 0);
99 const int32x4_t vksumzp4x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point45), 0);
100 const int32x4_t vksumzp5x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point45), 0);
101 const int32x4_t vinput_zero_point67 = vld1q_dup_s32(&quantization_params[6].zero_point);
102 const int32x4_t vksumzp6x0123 = vmulq_s32(vksum0123, vinput_zero_point67);
103 const int32x4_t vksumzp6x4567 = vmulq_s32(vksum4567, vinput_zero_point67);
104
105 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
106 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
107 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
108 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
109 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
110 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
111 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
112 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
113 int32x4_t vacc45x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
114 int32x4_t vacc45x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
115 int32x4_t vacc45x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
116 int32x4_t vacc45x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
117 int32x4_t vacc67x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp6x0123), vreinterpretq_u64_s32(vksumzp6x0123)));
118 int32x4_t vacc67x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp6x0123), vreinterpretq_u64_s32(vksumzp6x0123)));
119 int32x4_t vacc67x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp6x4567), vreinterpretq_u64_s32(vksumzp6x4567)));
120 int32x4_t vacc67x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp6x4567), vreinterpretq_u64_s32(vksumzp6x4567)));
121
122 // Inner accumulation loop along the 8 columns.
123 size_t k = kc;
124 // 2x partial unrolled loop to load 8 bytes at a time.
125
126 uint64x2x2_t va01x0123456789ABCDEF;
127 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
128 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
129 uint64x2x2_t va23x0123456789ABCDEF;
130 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
131 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
132 uint64x2x2_t va45x0123456789ABCDEF;
133 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
134 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
135 uint64x2x2_t va67x0123456789ABCDEF;
136 va67x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
137 va67x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
138
139 while (k >= 16 * sizeof(int8_t)) {
140 // Load a 7x16 block of activations.
141 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
142 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
143 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
144 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a6, va67x0123456789ABCDEF, 0); a6 += 16;
145 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
146 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
147 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
148
149 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
150 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
151 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
152 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
153 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
154 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
155 const int8x16_t va67x01234567 = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[0]);
156 const int8x16_t va67x89ABCDEF = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[1]);
157
158 // Load a 16x8 block of weights.
159 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
160 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
161 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
162 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
163 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
164 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
165 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
166 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
167
168 // Multiply-accumulate: 7x8 * 8x8 --> 7x8.
169 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
170 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
171 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
172 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
173 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
174 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
175 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
176 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
177 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
178 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
179 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
180 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
181 vacc67x01 = vmmlaq_s32(vacc67x01, va67x01234567, vb01x01234567);
182 vacc67x23 = vmmlaq_s32(vacc67x23, va67x01234567, vb23x01234567);
183 vacc67x45 = vmmlaq_s32(vacc67x45, va67x01234567, vb45x01234567);
184 vacc67x67 = vmmlaq_s32(vacc67x67, va67x01234567, vb67x01234567);
185 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
186 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
187 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
188 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
189 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
190 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
191 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
192 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
193 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
194 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
195 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
196 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
197 vacc67x01 = vmmlaq_s32(vacc67x01, va67x89ABCDEF, vb01x89ABCDEF);
198 vacc67x23 = vmmlaq_s32(vacc67x23, va67x89ABCDEF, vb23x89ABCDEF);
199 vacc67x45 = vmmlaq_s32(vacc67x45, va67x89ABCDEF, vb45x89ABCDEF);
200 vacc67x67 = vmmlaq_s32(vacc67x67, va67x89ABCDEF, vb67x89ABCDEF);
201
202 k -= 16 * sizeof(int8_t);
203 }
204 // Handle up to 8 final positions of `k`
205 if XNN_UNLIKELY(k != 0) {
206 // Load a 7x8 block of activations.
207 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
208 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
209 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
210 uint64x2_t va67x01234567 = vld1q_dup_u64((const void*) a6); a6 += 8;
211 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
212 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
213 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
214
215 // Load a 16x8 block of weights.
216 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
217 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
218 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
219 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
220
221 // Multiply-accumulate: 7x4 * 4x8 --> 7x8.
222 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
223 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
224 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
225 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
226 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
227 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
228 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
229 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
230 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
231 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
232 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
233 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
234 vacc67x01 = vmmlaq_s32(vacc67x01, vreinterpretq_s8_u64(va67x01234567), vb01x01234567);
235 vacc67x23 = vmmlaq_s32(vacc67x23, vreinterpretq_s8_u64(va67x01234567), vb23x01234567);
236 vacc67x45 = vmmlaq_s32(vacc67x45, vreinterpretq_s8_u64(va67x01234567), vb45x01234567);
237 vacc67x67 = vmmlaq_s32(vacc67x67, vreinterpretq_s8_u64(va67x01234567), vb67x01234567);
238 }
239
240 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
241 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
242 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
243 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
244 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
245 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
246 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
247 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
248 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
249 int32x4_t vacc5x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
250 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
251 int32x4_t vacc5x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
252 int32x4_t vacc6x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
253 int32x4_t vacc6x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
254 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
255 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
256 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
257 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
258 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
259 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
260 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
261 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
262 float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123);
263 float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567);
264 float32x4_t vout5x0123 = vcvtq_f32_s32(vacc5x0123);
265 float32x4_t vout5x4567 = vcvtq_f32_s32(vacc5x4567);
266 float32x4_t vout6x0123 = vcvtq_f32_s32(vacc6x0123);
267 float32x4_t vout6x4567 = vcvtq_f32_s32(vacc6x4567);
268
269 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
270 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
271 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
272 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
273 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
274 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
275 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
276 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
277 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
278 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
279 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
280 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
281 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
282 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
283 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
284 const float32x4_t vinput_scale6 = vld1q_dup_f32(&quantization_params[6].inv_scale);
285 vout6x0123 = vmulq_f32(vout6x0123, vinput_scale6);
286 vout6x4567 = vmulq_f32(vout6x4567, vinput_scale6);
287
288 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
289 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
290
291 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
292 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
293 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
294 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
295 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
296 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
297 vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
298 vout6x0123 = vfmaq_f32(vbias0123, vout6x0123, vfilter_output_scale0123);
299 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
300 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
301 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
302 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
303 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
304 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
305 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
306 vout6x4567 = vfmaq_f32(vbias4567, vout6x4567, vfilter_output_scale4567);
307
308 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
309 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
310 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
311 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
312 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
313 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
314 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
315 vout3x0123 = vmaxq_f32(vout3x0123, voutput_min);
316 vout3x4567 = vmaxq_f32(vout3x4567, voutput_min);
317 vout4x0123 = vmaxq_f32(vout4x0123, voutput_min);
318 vout4x4567 = vmaxq_f32(vout4x4567, voutput_min);
319 vout5x0123 = vmaxq_f32(vout5x0123, voutput_min);
320 vout5x4567 = vmaxq_f32(vout5x4567, voutput_min);
321 vout6x0123 = vmaxq_f32(vout6x0123, voutput_min);
322 vout6x4567 = vmaxq_f32(vout6x4567, voutput_min);
323
324 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
325 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
326 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
327 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
328 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
329 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
330 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
331 vout3x0123 = vminq_f32(vout3x0123, voutput_max);
332 vout3x4567 = vminq_f32(vout3x4567, voutput_max);
333 vout4x0123 = vminq_f32(vout4x0123, voutput_max);
334 vout4x4567 = vminq_f32(vout4x4567, voutput_max);
335 vout5x0123 = vminq_f32(vout5x0123, voutput_max);
336 vout5x4567 = vminq_f32(vout5x4567, voutput_max);
337 vout6x0123 = vminq_f32(vout6x0123, voutput_max);
338 vout6x4567 = vminq_f32(vout6x4567, voutput_max);
339
340 if XNN_LIKELY(nc >= 8) {
341 vst1q_f32(c0, vout0x0123);
342 vst1q_f32(c0 + 4, vout0x4567);
343 vst1q_f32(c1, vout1x0123);
344 vst1q_f32(c1 + 4, vout1x4567);
345 vst1q_f32(c2, vout2x0123);
346 vst1q_f32(c2 + 4, vout2x4567);
347 vst1q_f32(c3, vout3x0123);
348 vst1q_f32(c3 + 4, vout3x4567);
349 vst1q_f32(c4, vout4x0123);
350 vst1q_f32(c4 + 4, vout4x4567);
351 vst1q_f32(c5, vout5x0123);
352 vst1q_f32(c5 + 4, vout5x4567);
353 vst1q_f32(c6, vout6x0123);
354 vst1q_f32(c6 + 4, vout6x4567);
355
356 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
357 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
358 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
359 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
360 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
361 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
362 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
363
364 c0 = (float*) ((uintptr_t) c0 + cn_stride);
365 c1 = (float*) ((uintptr_t) c1 + cn_stride);
366 c2 = (float*) ((uintptr_t) c2 + cn_stride);
367 c3 = (float*) ((uintptr_t) c3 + cn_stride);
368 c4 = (float*) ((uintptr_t) c4 + cn_stride);
369 c5 = (float*) ((uintptr_t) c5 + cn_stride);
370 c6 = (float*) ((uintptr_t) c6 + cn_stride);
371
372 nc -= 8;
373 } else {
374 if (nc & 4) {
375 vst1q_f32(c0, vout0x0123); c0 += 4;
376 vout0x0123 = vout0x4567;
377 vst1q_f32(c1, vout1x0123); c1 += 4;
378 vout1x0123 = vout1x4567;
379 vst1q_f32(c2, vout2x0123); c2 += 4;
380 vout2x0123 = vout2x4567;
381 vst1q_f32(c3, vout3x0123); c3 += 4;
382 vout3x0123 = vout3x4567;
383 vst1q_f32(c4, vout4x0123); c4 += 4;
384 vout4x0123 = vout4x4567;
385 vst1q_f32(c5, vout5x0123); c5 += 4;
386 vout5x0123 = vout5x4567;
387 vst1q_f32(c6, vout6x0123); c6 += 4;
388 vout6x0123 = vout6x4567;
389 }
390 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
391 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
392 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
393 float32x2_t vout3x01 = vget_low_f32(vout3x0123);
394 float32x2_t vout4x01 = vget_low_f32(vout4x0123);
395 float32x2_t vout5x01 = vget_low_f32(vout5x0123);
396 float32x2_t vout6x01 = vget_low_f32(vout6x0123);
397 if (nc & 2) {
398 vst1_f32(c0, vout0x01); c0 += 2;
399 vst1_f32(c1, vout1x01); c1 += 2;
400 vst1_f32(c2, vout2x01); c2 += 2;
401 vst1_f32(c3, vout3x01); c3 += 2;
402 vst1_f32(c4, vout4x01); c4 += 2;
403 vst1_f32(c5, vout5x01); c5 += 2;
404 vst1_f32(c6, vout6x01); c6 += 2;
405 vout0x01 = vget_high_f32(vout0x0123);
406 vout1x01 = vget_high_f32(vout1x0123);
407 vout2x01 = vget_high_f32(vout2x0123);
408 vout3x01 = vget_high_f32(vout3x0123);
409 vout4x01 = vget_high_f32(vout4x0123);
410 vout5x01 = vget_high_f32(vout5x0123);
411 vout6x01 = vget_high_f32(vout6x0123);
412 }
413 if (nc & 1) {
414 vst1_lane_f32(c0, vout0x01, 0);
415 vst1_lane_f32(c1, vout1x01, 0);
416 vst1_lane_f32(c2, vout2x01, 0);
417 vst1_lane_f32(c3, vout3x01, 0);
418 vst1_lane_f32(c4, vout4x01, 0);
419 vst1_lane_f32(c5, vout5x01, 0);
420 vst1_lane_f32(c6, vout6x01, 0);
421 }
422 nc = 0;
423 }
424 } while (nc != 0);
425 }