Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-8x8c8-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_8x8c8__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 <= 8);
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 int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr < 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
66 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 4) {
68 a4 = a3;
69 c4 = c3;
70 }
71 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
72 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 6) {
74 a5 = a4;
75 c5 = c4;
76 }
77 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
78 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
79 if XNN_UNPREDICTABLE(mr <= 6) {
80 a6 = a5;
81 c6 = c5;
82 }
83 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
84 uint16_t* c7 = (uint16_t*) ((uintptr_t) c6 + cm_stride);
85 if XNN_UNPREDICTABLE(mr != 8) {
86 a7 = a6;
87 c7 = c6;
88 }
89 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
90
91 // Loop over groups of 8 columns.
92 do {
93 // Initialize accumulators with scaled vksum. 8 scaled vksum values are loaded from the
94 // weight matrix, at the start of the group of 8 columns.
95 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
96 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
97 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
98 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
99 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
100 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
101 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
102 const float32x4_t vinput_zero_point23 = vcvtq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
103 float32x4_t vout2x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point23), 0);
104 float32x4_t vout3x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point23), 0);
105 float32x4_t vout2x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point23), 0);
106 float32x4_t vout3x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point23), 0);
107 const float32x4_t vinput_zero_point45 = vcvtq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
108 float32x4_t vout4x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point45), 0);
109 float32x4_t vout5x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point45), 0);
110 float32x4_t vout4x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point45), 0);
111 float32x4_t vout5x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point45), 0);
112 const float32x4_t vinput_zero_point67 = vcvtq_f32_s32(vld1q_s32(&quantization_params[6].zero_point));
113 float32x4_t vout6x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point67), 0);
114 float32x4_t vout7x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point67), 0);
115 float32x4_t vout6x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point67), 0);
116 float32x4_t vout7x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point67), 0);
117
118
119 for (size_t kb=0; kb < kc; kb += bl) {
120 int32x4_t vacc01x01 = vdupq_n_s32(0);
121 int32x4_t vacc01x23 = vdupq_n_s32(0);
122 int32x4_t vacc01x45 = vdupq_n_s32(0);
123 int32x4_t vacc01x67 = vdupq_n_s32(0);
124 int32x4_t vacc23x01 = vdupq_n_s32(0);
125 int32x4_t vacc23x23 = vdupq_n_s32(0);
126 int32x4_t vacc23x45 = vdupq_n_s32(0);
127 int32x4_t vacc23x67 = vdupq_n_s32(0);
128 int32x4_t vacc45x01 = vdupq_n_s32(0);
129 int32x4_t vacc45x23 = vdupq_n_s32(0);
130 int32x4_t vacc45x45 = vdupq_n_s32(0);
131 int32x4_t vacc45x67 = vdupq_n_s32(0);
132 int32x4_t vacc67x01 = vdupq_n_s32(0);
133 int32x4_t vacc67x23 = vdupq_n_s32(0);
134 int32x4_t vacc67x45 = vdupq_n_s32(0);
135 int32x4_t vacc67x67 = vdupq_n_s32(0);
136
137 size_t k = bl;
138 // 2x partial unrolled loop to load 8 bytes at a time.
139
140 uint64x2x2_t va01x0123456789ABCDEF;
141 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
142 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
143 uint64x2x2_t va23x0123456789ABCDEF;
144 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
145 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
146 uint64x2x2_t va45x0123456789ABCDEF;
147 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
148 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
149 uint64x2x2_t va67x0123456789ABCDEF;
150 va67x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
151 va67x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
152
153 while (k >= 16 * sizeof(int8_t)) {
154 // Load a 8x16 block of activations.
155 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
156 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
157 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
158 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a6, va67x0123456789ABCDEF, 0); a6 += 16;
159 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
160 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
161 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
162 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a7, va67x0123456789ABCDEF, 1); a7 += 16;
163
164 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
165 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
166 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
167 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
168 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
169 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
170 const int8x16_t va67x01234567 = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[0]);
171 const int8x16_t va67x89ABCDEF = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[1]);
172
173 // Load a 16x8 block of weights.
174 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
175 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
176 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
177 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
178 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
179 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
180 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
181 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
182 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
183 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
184 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
185 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
186
187 // Multiply-accumulate: 8x8 * 8x8 --> 8x8.
188 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
189 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
190 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
191 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
192 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
193 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
194 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
195 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
196 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
197 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
198 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
199 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
200 vacc67x01 = vmmlaq_s32(vacc67x01, va67x01234567, vb01x01234567);
201 vacc67x23 = vmmlaq_s32(vacc67x23, va67x01234567, vb23x01234567);
202 vacc67x45 = vmmlaq_s32(vacc67x45, va67x01234567, vb45x01234567);
203 vacc67x67 = vmmlaq_s32(vacc67x67, va67x01234567, vb67x01234567);
204 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
205 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
206 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
207 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
208 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
209 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
210 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
211 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
212 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
213 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
214 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
215 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
216 vacc67x01 = vmmlaq_s32(vacc67x01, va67x89ABCDEF, vb01x89ABCDEF);
217 vacc67x23 = vmmlaq_s32(vacc67x23, va67x89ABCDEF, vb23x89ABCDEF);
218 vacc67x45 = vmmlaq_s32(vacc67x45, va67x89ABCDEF, vb45x89ABCDEF);
219 vacc67x67 = vmmlaq_s32(vacc67x67, va67x89ABCDEF, vb67x89ABCDEF);
220
221 k -= 16 * sizeof(int8_t);
222 }
223 // Handle up to 8 final positions of `k`
224 if XNN_UNLIKELY(k != 0) {
225 // Load a 8x8 block of activations.
226 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
227 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
228 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
229 uint64x2_t va67x01234567 = vld1q_dup_u64((const void*) a6); a6 += 8;
230 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
231 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
232 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
233 va67x01234567 = vld1q_lane_u64((const void*) a7, va67x01234567, 1); a7 += 8;
234
235 // Load a 16x8 block of weights.
236 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
237 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
238 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
239 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
240 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
241 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
242 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
243 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
244
245 // Multiply-accumulate: 8x4 * 4x8 --> 8x8.
246 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
247 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
248 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
249 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
250 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
251 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
252 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
253 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
254 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
255 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
256 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
257 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
258 vacc67x01 = vmmlaq_s32(vacc67x01, vreinterpretq_s8_u64(va67x01234567), vb01x01234567);
259 vacc67x23 = vmmlaq_s32(vacc67x23, vreinterpretq_s8_u64(va67x01234567), vb23x01234567);
260 vacc67x45 = vmmlaq_s32(vacc67x45, vreinterpretq_s8_u64(va67x01234567), vb45x01234567);
261 vacc67x67 = vmmlaq_s32(vacc67x67, vreinterpretq_s8_u64(va67x01234567), vb67x01234567);
262 }
263
264 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
265 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
266 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
267 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
268 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
269 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
270 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
271 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
272 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
273 int32x4_t vacc5x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
274 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
275 int32x4_t vacc5x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
276 int32x4_t vacc6x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
277 int32x4_t vacc7x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
278 int32x4_t vacc6x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
279 int32x4_t vacc7x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
280 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
281 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
282 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
283 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
284 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
285 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
286 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
287 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
288 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
289 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
290 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
291 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
292 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
293 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
294 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
295 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
296 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
297 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
298 float32x4_t vf4x0123 = vcvtq_f32_s32(vacc4x0123);
299 vout4x0123 = vfmaq_f32(vout4x0123, vf4x0123, vfilter_output_scale0123);
300 float32x4_t vf4x4567 = vcvtq_f32_s32(vacc4x4567);
301 vout4x4567 = vfmaq_f32(vout4x4567, vf4x4567, vfilter_output_scale4567);
302 float32x4_t vf5x0123 = vcvtq_f32_s32(vacc5x0123);
303 vout5x0123 = vfmaq_f32(vout5x0123, vf5x0123, vfilter_output_scale0123);
304 float32x4_t vf5x4567 = vcvtq_f32_s32(vacc5x4567);
305 vout5x4567 = vfmaq_f32(vout5x4567, vf5x4567, vfilter_output_scale4567);
306 float32x4_t vf6x0123 = vcvtq_f32_s32(vacc6x0123);
307 vout6x0123 = vfmaq_f32(vout6x0123, vf6x0123, vfilter_output_scale0123);
308 float32x4_t vf6x4567 = vcvtq_f32_s32(vacc6x4567);
309 vout6x4567 = vfmaq_f32(vout6x4567, vf6x4567, vfilter_output_scale4567);
310 float32x4_t vf7x0123 = vcvtq_f32_s32(vacc7x0123);
311 vout7x0123 = vfmaq_f32(vout7x0123, vf7x0123, vfilter_output_scale0123);
312 float32x4_t vf7x4567 = vcvtq_f32_s32(vacc7x4567);
313 vout7x4567 = vfmaq_f32(vout7x4567, vf7x4567, vfilter_output_scale4567);
314 }
315
316 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
317 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
318 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
319 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
320 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
321 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
322 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
323 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
324 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
325 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
326 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
327 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
328 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
329 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
330 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
331 const float32x4_t vinput_scale67 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[6].zero_point));
332 vout6x0123 = vmulq_lane_f32(vout6x0123, vget_low_f32(vinput_scale67), 1);
333 vout7x0123 = vmulq_lane_f32(vout7x0123, vget_high_f32(vinput_scale67), 1);
334 vout6x4567 = vmulq_lane_f32(vout6x4567, vget_low_f32(vinput_scale67), 1);
335 vout7x4567 = vmulq_lane_f32(vout7x4567, vget_high_f32(vinput_scale67), 1);
336
337
338 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
339 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
340 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
341 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
342 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
343 vout4x0123 = vaddq_f32(vbias0123, vout4x0123);
344 vout5x0123 = vaddq_f32(vbias0123, vout5x0123);
345 vout6x0123 = vaddq_f32(vbias0123, vout6x0123);
346 vout7x0123 = vaddq_f32(vbias0123, vout7x0123);
347 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
348 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
349 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
350 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
351 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
352 vout4x4567 = vaddq_f32(vbias4567, vout4x4567);
353 vout5x4567 = vaddq_f32(vbias4567, vout5x4567);
354 vout6x4567 = vaddq_f32(vbias4567, vout6x4567);
355 vout7x4567 = vaddq_f32(vbias4567, vout7x4567);
356
357 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
358 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
359 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
360 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
361 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
362 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
363 float16x8_t vfp16out6x01234567 = vcombine_f16(vcvt_f16_f32(vout6x0123), vcvt_f16_f32(vout6x4567));
364 float16x8_t vfp16out7x01234567 = vcombine_f16(vcvt_f16_f32(vout7x0123), vcvt_f16_f32(vout7x4567));
365
366 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
367 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
368 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
369 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
370 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
371 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
372 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
373 vfp16out6x01234567 = vmaxq_f16(vfp16out6x01234567, voutput_min);
374 vfp16out7x01234567 = vmaxq_f16(vfp16out7x01234567, 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 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
378 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
379 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
380 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
381 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
382 vfp16out6x01234567 = vminq_f16(vfp16out6x01234567, voutput_max);
383 vfp16out7x01234567 = vminq_f16(vfp16out7x01234567, voutput_max);
384 if XNN_LIKELY(nc >= 8) {
385 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
386 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
387 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
388 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
389 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
390 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
391 vst1q_u16(c6, vreinterpretq_u16_f16(vfp16out6x01234567));
392 vst1q_u16(c7, vreinterpretq_u16_f16(vfp16out7x01234567));
393
394 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
395 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
396 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
397 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
398 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
399 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
400 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
401 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
402
403 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
404 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
405 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
406 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
407 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
408 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
409 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
410 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
411
412 nc -= 8;
413 } else {
414 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
415 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
416 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
417 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
418 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
419 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
420 float16x4_t vfp16out6x0123 = vget_low_f16(vfp16out6x01234567);
421 float16x4_t vfp16out7x0123 = vget_low_f16(vfp16out7x01234567);
422 if (nc & 4) {
423 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
424 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
425 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
426 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
427 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
428 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
429 vst1_u16(c6, vreinterpret_u16_f16(vfp16out6x0123)); c6 += 4;
430 vst1_u16(c7, vreinterpret_u16_f16(vfp16out7x0123)); c7 += 4;
431 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
432 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
433 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
434 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
435 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
436 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
437 vfp16out6x0123 = vget_high_f16(vfp16out6x01234567);
438 vfp16out7x0123 = vget_high_f16(vfp16out7x01234567);
439 }
440 if (nc & 2) {
441 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
442 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
443 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
444 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
445 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
446 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
447 vst1_lane_u32((void*) c6, vreinterpret_u32_f16(vfp16out6x0123), 0); c6 += 2;
448 vst1_lane_u32((void*) c7, vreinterpret_u32_f16(vfp16out7x0123), 0); c7 += 2;
449 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
450 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
451 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
452 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
453 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
454 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
455 vfp16out6x0123 = vext_f16(vfp16out6x0123, vfp16out6x0123, 2);
456 vfp16out7x0123 = vext_f16(vfp16out7x0123, vfp16out7x0123, 2);
457 }
458 if (nc & 1) {
459 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
460 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
461 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
462 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
463 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
464 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
465 vst1_lane_u16(c6, vreinterpret_u16_f16(vfp16out6x0123), 0);
466 vst1_lane_u16(c7, vreinterpret_u16_f16(vfp16out7x0123), 0);
467 }
468 nc = 0;
469 }
470 } while (nc != 0);
471 }