Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-1x32c8-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_f16_qb4w_gemm_minmax_ukernel_1x32c8__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 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const struct xnn_f16_qb4w_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 <= 1);
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 size_t bl = params->scalar.blocksize;
42 assert(bl <= kc);
43 assert(bl != 0);
44 assert(bl % 32 == 0);
45 const int8_t* a0 = a;
46 uint16_t* c0 = (uint16_t*) c;
47 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
48
49 // Loop over groups of 32 columns.
50 do {
51 // Initialize accumulators with scaled vksum. 32 scaled vksum values are loaded from the
52 // weight matrix, at the start of the group of 32 columns.
53 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[0].zero_point));
54 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
55 float32x4_t vout0x0123 = vmulq_f32(vksum0123, vinput_zero_point01);
56 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
57 float32x4_t vout0x4567 = vmulq_f32(vksum4567, vinput_zero_point01);
58 const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
59 float32x4_t vout0x89AB = vmulq_f32(vksum89AB, vinput_zero_point01);
60 const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
61 float32x4_t vout0xCDEF = vmulq_f32(vksumCDEF, vinput_zero_point01);
62 const float32x4_t vksumGHIJ = vld1q_f32(w); w = (const float*) w + 4;
63 float32x4_t vout0xGHIJ = vmulq_f32(vksumGHIJ, vinput_zero_point01);
64 const float32x4_t vksumKLMN = vld1q_f32(w); w = (const float*) w + 4;
65 float32x4_t vout0xKLMN = vmulq_f32(vksumKLMN, vinput_zero_point01);
66 const float32x4_t vksumOPQR = vld1q_f32(w); w = (const float*) w + 4;
67 float32x4_t vout0xOPQR = vmulq_f32(vksumOPQR, vinput_zero_point01);
68 const float32x4_t vksumSTUV = vld1q_f32(w); w = (const float*) w + 4;
69 float32x4_t vout0xSTUV = vmulq_f32(vksumSTUV, vinput_zero_point01);
70
71
72 for (size_t kb=0; kb < kc; kb += bl) {
73 int32x4_t vacc01x01 = vdupq_n_s32(0);
74 int32x4_t vacc01x23 = vdupq_n_s32(0);
75 int32x4_t vacc01x45 = vdupq_n_s32(0);
76 int32x4_t vacc01x67 = vdupq_n_s32(0);
77 int32x4_t vacc01x89 = vdupq_n_s32(0);
78 int32x4_t vacc01xAB = vdupq_n_s32(0);
79 int32x4_t vacc01xCD = vdupq_n_s32(0);
80 int32x4_t vacc01xEF = vdupq_n_s32(0);
81 int32x4_t vacc01xGH = vdupq_n_s32(0);
82 int32x4_t vacc01xIJ = vdupq_n_s32(0);
83 int32x4_t vacc01xKL = vdupq_n_s32(0);
84 int32x4_t vacc01xMN = vdupq_n_s32(0);
85 int32x4_t vacc01xOP = vdupq_n_s32(0);
86 int32x4_t vacc01xQR = vdupq_n_s32(0);
87 int32x4_t vacc01xST = vdupq_n_s32(0);
88 int32x4_t vacc01xUV = vdupq_n_s32(0);
89
90 size_t k = bl;
91 // 2x partial unrolled loop to load 8 bytes at a time.
92
93 uint64x2x2_t va01x0123456789ABCDEF;
94 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
95 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
96
97 while (k >= 16 * sizeof(int8_t)) {
98 // Load a 1x16 block of activations.
99 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
100
101 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
102 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
103
104 // Load a 16x32 block of weights.
105 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
106 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
107 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
108 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
109 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
110 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
111 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
112 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
113 const int8x16_t vbGHx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
114 const int8x16_t vbIJx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
115 const int8x16_t vbKLx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
116 const int8x16_t vbMNx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
117 const int8x16_t vbOPx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
118 const int8x16_t vbQRx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
119 const int8x16_t vbSTx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
120 const int8x16_t vbUVx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
121 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
122 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
123 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
124 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
125 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
126 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
127 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
128 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
129 const int8x16_t vbGHx01234567 = vshlq_n_s8(vbGHx0123456789ABCDEF, 4);
130 const int8x16_t vbIJx01234567 = vshlq_n_s8(vbIJx0123456789ABCDEF, 4);
131 const int8x16_t vbKLx01234567 = vshlq_n_s8(vbKLx0123456789ABCDEF, 4);
132 const int8x16_t vbMNx01234567 = vshlq_n_s8(vbMNx0123456789ABCDEF, 4);
133 const int8x16_t vbOPx01234567 = vshlq_n_s8(vbOPx0123456789ABCDEF, 4);
134 const int8x16_t vbQRx01234567 = vshlq_n_s8(vbQRx0123456789ABCDEF, 4);
135 const int8x16_t vbSTx01234567 = vshlq_n_s8(vbSTx0123456789ABCDEF, 4);
136 const int8x16_t vbUVx01234567 = vshlq_n_s8(vbUVx0123456789ABCDEF, 4);
137 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
138 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
139 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
140 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
141 const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask);
142 const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask);
143 const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask);
144 const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask);
145 const int8x16_t vbGHx89ABCDEF = vandq_s8(vbGHx0123456789ABCDEF, vmask);
146 const int8x16_t vbIJx89ABCDEF = vandq_s8(vbIJx0123456789ABCDEF, vmask);
147 const int8x16_t vbKLx89ABCDEF = vandq_s8(vbKLx0123456789ABCDEF, vmask);
148 const int8x16_t vbMNx89ABCDEF = vandq_s8(vbMNx0123456789ABCDEF, vmask);
149 const int8x16_t vbOPx89ABCDEF = vandq_s8(vbOPx0123456789ABCDEF, vmask);
150 const int8x16_t vbQRx89ABCDEF = vandq_s8(vbQRx0123456789ABCDEF, vmask);
151 const int8x16_t vbSTx89ABCDEF = vandq_s8(vbSTx0123456789ABCDEF, vmask);
152 const int8x16_t vbUVx89ABCDEF = vandq_s8(vbUVx0123456789ABCDEF, vmask);
153
154 // Multiply-accumulate: 1x8 * 8x32 --> 1x32.
155 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
156 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
157 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
158 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
159 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
160 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
161 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
162 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
163 vacc01xGH = vmmlaq_s32(vacc01xGH, va01x01234567, vbGHx01234567);
164 vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x01234567, vbIJx01234567);
165 vacc01xKL = vmmlaq_s32(vacc01xKL, va01x01234567, vbKLx01234567);
166 vacc01xMN = vmmlaq_s32(vacc01xMN, va01x01234567, vbMNx01234567);
167 vacc01xOP = vmmlaq_s32(vacc01xOP, va01x01234567, vbOPx01234567);
168 vacc01xQR = vmmlaq_s32(vacc01xQR, va01x01234567, vbQRx01234567);
169 vacc01xST = vmmlaq_s32(vacc01xST, va01x01234567, vbSTx01234567);
170 vacc01xUV = vmmlaq_s32(vacc01xUV, va01x01234567, vbUVx01234567);
171 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
172 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
173 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
174 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
175 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
176 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
177 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
178 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
179 vacc01xGH = vmmlaq_s32(vacc01xGH, va01x89ABCDEF, vbGHx89ABCDEF);
180 vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x89ABCDEF, vbIJx89ABCDEF);
181 vacc01xKL = vmmlaq_s32(vacc01xKL, va01x89ABCDEF, vbKLx89ABCDEF);
182 vacc01xMN = vmmlaq_s32(vacc01xMN, va01x89ABCDEF, vbMNx89ABCDEF);
183 vacc01xOP = vmmlaq_s32(vacc01xOP, va01x89ABCDEF, vbOPx89ABCDEF);
184 vacc01xQR = vmmlaq_s32(vacc01xQR, va01x89ABCDEF, vbQRx89ABCDEF);
185 vacc01xST = vmmlaq_s32(vacc01xST, va01x89ABCDEF, vbSTx89ABCDEF);
186 vacc01xUV = vmmlaq_s32(vacc01xUV, va01x89ABCDEF, vbUVx89ABCDEF);
187
188 k -= 16 * sizeof(int8_t);
189 }
190 // Handle up to 8 final positions of `k`
191 if XNN_UNLIKELY(k != 0) {
192 // Load a 1x8 block of activations.
193 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
194
195 // Load a 16x32 block of weights.
196 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
197 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
198 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
199 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
200 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
201 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
202 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
203 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
204 const int8x16_t vbGHx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
205 const int8x16_t vbIJx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
206 const int8x16_t vbKLx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
207 const int8x16_t vbMNx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
208 const int8x16_t vbOPx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
209 const int8x16_t vbQRx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
210 const int8x16_t vbSTx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
211 const int8x16_t vbUVx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
212 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
213 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
214 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
215 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
216 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
217 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
218 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
219 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
220 const int8x16_t vbGHx01234567 = vshlq_n_s8(vbGHx0123456789ABCDEF, 4);
221 const int8x16_t vbIJx01234567 = vshlq_n_s8(vbIJx0123456789ABCDEF, 4);
222 const int8x16_t vbKLx01234567 = vshlq_n_s8(vbKLx0123456789ABCDEF, 4);
223 const int8x16_t vbMNx01234567 = vshlq_n_s8(vbMNx0123456789ABCDEF, 4);
224 const int8x16_t vbOPx01234567 = vshlq_n_s8(vbOPx0123456789ABCDEF, 4);
225 const int8x16_t vbQRx01234567 = vshlq_n_s8(vbQRx0123456789ABCDEF, 4);
226 const int8x16_t vbSTx01234567 = vshlq_n_s8(vbSTx0123456789ABCDEF, 4);
227 const int8x16_t vbUVx01234567 = vshlq_n_s8(vbUVx0123456789ABCDEF, 4);
228
229 // Multiply-accumulate: 1x4 * 4x32 --> 1x32.
230 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
231 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
232 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
233 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
234 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
235 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
236 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
237 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
238 vacc01xGH = vmmlaq_s32(vacc01xGH, vreinterpretq_s8_u64(va01x01234567), vbGHx01234567);
239 vacc01xIJ = vmmlaq_s32(vacc01xIJ, vreinterpretq_s8_u64(va01x01234567), vbIJx01234567);
240 vacc01xKL = vmmlaq_s32(vacc01xKL, vreinterpretq_s8_u64(va01x01234567), vbKLx01234567);
241 vacc01xMN = vmmlaq_s32(vacc01xMN, vreinterpretq_s8_u64(va01x01234567), vbMNx01234567);
242 vacc01xOP = vmmlaq_s32(vacc01xOP, vreinterpretq_s8_u64(va01x01234567), vbOPx01234567);
243 vacc01xQR = vmmlaq_s32(vacc01xQR, vreinterpretq_s8_u64(va01x01234567), vbQRx01234567);
244 vacc01xST = vmmlaq_s32(vacc01xST, vreinterpretq_s8_u64(va01x01234567), vbSTx01234567);
245 vacc01xUV = vmmlaq_s32(vacc01xUV, vreinterpretq_s8_u64(va01x01234567), vbUVx01234567);
246 }
247
248 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
249 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
250 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
251 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
252 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xGH), vreinterpretq_u64_s32(vacc01xIJ)));
253 int32x4_t vacc0xKLMN = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xKL), vreinterpretq_u64_s32(vacc01xMN)));
254 int32x4_t vacc0xOPQR = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xOP), vreinterpretq_u64_s32(vacc01xQR)));
255 int32x4_t vacc0xSTUV = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xST), vreinterpretq_u64_s32(vacc01xUV)));
256 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
257 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
258 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
259 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
260 const float32x4_t vfilter_output_scaleGHIJ = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
261 const float32x4_t vfilter_output_scaleKLMN = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
262 const float32x4_t vfilter_output_scaleOPQR = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
263 const float32x4_t vfilter_output_scaleSTUV = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
264 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
265 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
266 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
267 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
268 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
269 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
270 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
271 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
272 float32x4_t vf0xGHIJ = vcvtq_f32_s32(vacc0xGHIJ);
273 vout0xGHIJ = vfmaq_f32(vout0xGHIJ, vf0xGHIJ, vfilter_output_scaleGHIJ);
274 float32x4_t vf0xKLMN = vcvtq_f32_s32(vacc0xKLMN);
275 vout0xKLMN = vfmaq_f32(vout0xKLMN, vf0xKLMN, vfilter_output_scaleKLMN);
276 float32x4_t vf0xOPQR = vcvtq_f32_s32(vacc0xOPQR);
277 vout0xOPQR = vfmaq_f32(vout0xOPQR, vf0xOPQR, vfilter_output_scaleOPQR);
278 float32x4_t vf0xSTUV = vcvtq_f32_s32(vacc0xSTUV);
279 vout0xSTUV = vfmaq_f32(vout0xSTUV, vf0xSTUV, vfilter_output_scaleSTUV);
280 }
281
282 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
283 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
284 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
285 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
286 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
287 vout0xGHIJ = vmulq_f32(vout0xGHIJ, vinput_scale0);
288 vout0xKLMN = vmulq_f32(vout0xKLMN, vinput_scale0);
289 vout0xOPQR = vmulq_f32(vout0xOPQR, vinput_scale0);
290 vout0xSTUV = vmulq_f32(vout0xSTUV, vinput_scale0);
291
292
293 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
294 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
295 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
296 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
297 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
298 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
299 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
300 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
301 const float32x4_t vbiasGHIJ = vld1q_f32(w); w = (const float*) w + 4;
302 vout0xGHIJ = vaddq_f32(vbiasGHIJ, vout0xGHIJ);
303 const float32x4_t vbiasKLMN = vld1q_f32(w); w = (const float*) w + 4;
304 vout0xKLMN = vaddq_f32(vbiasKLMN, vout0xKLMN);
305 const float32x4_t vbiasOPQR = vld1q_f32(w); w = (const float*) w + 4;
306 vout0xOPQR = vaddq_f32(vbiasOPQR, vout0xOPQR);
307 const float32x4_t vbiasSTUV = vld1q_f32(w); w = (const float*) w + 4;
308 vout0xSTUV = vaddq_f32(vbiasSTUV, vout0xSTUV);
309
310 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
311 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
312 float16x8_t vfp16out0xGHIJKLMN = vcombine_f16(vcvt_f16_f32(vout0xGHIJ), vcvt_f16_f32(vout0xKLMN));
313 float16x8_t vfp16out0xOPQRSTUV = vcombine_f16(vcvt_f16_f32(vout0xOPQR), vcvt_f16_f32(vout0xSTUV));
314
315 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
316 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
317 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
318 vfp16out0xGHIJKLMN = vmaxq_f16(vfp16out0xGHIJKLMN, voutput_min);
319 vfp16out0xOPQRSTUV = vmaxq_f16(vfp16out0xOPQRSTUV, voutput_min);
320 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
321 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
322 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
323 vfp16out0xGHIJKLMN = vminq_f16(vfp16out0xGHIJKLMN, voutput_max);
324 vfp16out0xOPQRSTUV = vminq_f16(vfp16out0xOPQRSTUV, voutput_max);
325 if XNN_LIKELY(nc >= 32) {
326 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
327 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
328 vst1q_u16(c0 + 16, vreinterpretq_u16_f16(vfp16out0xGHIJKLMN));
329 vst1q_u16(c0 + 24, vreinterpretq_u16_f16(vfp16out0xOPQRSTUV));
330
331 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
332
333 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
334
335 nc -= 32;
336 } else {
337 if (nc & 16) {
338 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
339 vfp16out0x01234567 = vfp16out0xGHIJKLMN;
340 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x89ABCDEF)); c0 += 8;
341 vfp16out0x89ABCDEF = vfp16out0xOPQRSTUV;
342 }
343 if (nc & 8) {
344 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
345 vfp16out0x01234567 = vfp16out0x89ABCDEF;
346 }
347 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
348 if (nc & 4) {
349 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
350 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
351 }
352 if (nc & 2) {
353 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
354 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
355 }
356 if (nc & 1) {
357 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
358 }
359 nc = 0;
360 }
361 } while (nc != 0);
362 }