Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-6x8c8-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_6x8c8__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 <= 6);
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 int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
78
79 // Loop over groups of 8 columns.
80 do {
81 // Initialize accumulators with scaled vksum. 8 scaled vksum values are loaded from the
82 // weight matrix, at the start of the group of 8 columns.
83 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
84 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
85 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
86 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
87 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
88 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
89 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
90 const float32x4_t vinput_zero_point23 = vcvtq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
91 float32x4_t vout2x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point23), 0);
92 float32x4_t vout3x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point23), 0);
93 float32x4_t vout2x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point23), 0);
94 float32x4_t vout3x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point23), 0);
95 const float32x4_t vinput_zero_point45 = vcvtq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
96 float32x4_t vout4x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point45), 0);
97 float32x4_t vout5x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point45), 0);
98 float32x4_t vout4x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point45), 0);
99 float32x4_t vout5x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point45), 0);
100
101
102 for (size_t kb=0; kb < kc; kb += bl) {
103 int32x4_t vacc01x01 = vdupq_n_s32(0);
104 int32x4_t vacc01x23 = vdupq_n_s32(0);
105 int32x4_t vacc01x45 = vdupq_n_s32(0);
106 int32x4_t vacc01x67 = vdupq_n_s32(0);
107 int32x4_t vacc23x01 = vdupq_n_s32(0);
108 int32x4_t vacc23x23 = vdupq_n_s32(0);
109 int32x4_t vacc23x45 = vdupq_n_s32(0);
110 int32x4_t vacc23x67 = vdupq_n_s32(0);
111 int32x4_t vacc45x01 = vdupq_n_s32(0);
112 int32x4_t vacc45x23 = vdupq_n_s32(0);
113 int32x4_t vacc45x45 = vdupq_n_s32(0);
114 int32x4_t vacc45x67 = vdupq_n_s32(0);
115
116 size_t k = bl;
117 // 2x partial unrolled loop to load 8 bytes at a time.
118
119 uint64x2x2_t va01x0123456789ABCDEF;
120 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
121 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
122 uint64x2x2_t va23x0123456789ABCDEF;
123 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
124 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
125 uint64x2x2_t va45x0123456789ABCDEF;
126 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
127 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
128
129 while (k >= 16 * sizeof(int8_t)) {
130 // Load a 6x16 block of activations.
131 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
132 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
133 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
134 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
135 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
136 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
137
138 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
139 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
140 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
141 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
142 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
143 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
144
145 // Load a 16x8 block of weights.
146 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
147 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
148 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
149 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
150 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
151 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
152 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
153 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
154 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
155 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
156 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
157 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
158
159 // Multiply-accumulate: 6x8 * 8x8 --> 6x8.
160 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
161 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
162 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
163 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
164 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
165 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
166 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
167 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
168 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
169 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
170 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
171 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
172 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
173 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
174 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
175 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
176 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
177 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
178 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
179 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
180 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
181 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
182 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
183 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
184
185 k -= 16 * sizeof(int8_t);
186 }
187 // Handle up to 8 final positions of `k`
188 if XNN_UNLIKELY(k != 0) {
189 // Load a 6x8 block of activations.
190 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
191 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
192 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
193 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
194 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
195 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
196
197 // Load a 16x8 block of weights.
198 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
199 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
200 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
201 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
202 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
203 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
204 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
205 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
206
207 // Multiply-accumulate: 6x4 * 4x8 --> 6x8.
208 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
209 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
210 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
211 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
212 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
213 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
214 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
215 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
216 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
217 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
218 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
219 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
220 }
221
222 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
223 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
224 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
225 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
226 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
227 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
228 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
229 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
230 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
231 int32x4_t vacc5x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
232 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
233 int32x4_t vacc5x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
234 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
235 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
236 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
237 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
238 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
239 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
240 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
241 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
242 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
243 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
244 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
245 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
246 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
247 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
248 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
249 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
250 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
251 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
252 float32x4_t vf4x0123 = vcvtq_f32_s32(vacc4x0123);
253 vout4x0123 = vfmaq_f32(vout4x0123, vf4x0123, vfilter_output_scale0123);
254 float32x4_t vf4x4567 = vcvtq_f32_s32(vacc4x4567);
255 vout4x4567 = vfmaq_f32(vout4x4567, vf4x4567, vfilter_output_scale4567);
256 float32x4_t vf5x0123 = vcvtq_f32_s32(vacc5x0123);
257 vout5x0123 = vfmaq_f32(vout5x0123, vf5x0123, vfilter_output_scale0123);
258 float32x4_t vf5x4567 = vcvtq_f32_s32(vacc5x4567);
259 vout5x4567 = vfmaq_f32(vout5x4567, vf5x4567, vfilter_output_scale4567);
260 }
261
262 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
263 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
264 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
265 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
266 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
267 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
268 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
269 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
270 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
271 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
272 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
273 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
274 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
275 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
276 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
277
278
279 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
280 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
281 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
282 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
283 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
284 vout4x0123 = vaddq_f32(vbias0123, vout4x0123);
285 vout5x0123 = vaddq_f32(vbias0123, vout5x0123);
286 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
287 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
288 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
289 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
290 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
291 vout4x4567 = vaddq_f32(vbias4567, vout4x4567);
292 vout5x4567 = vaddq_f32(vbias4567, vout5x4567);
293
294 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
295 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
296 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
297 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
298 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
299 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
300
301 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
302 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
303 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
304 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
305 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
306 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
307 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
308 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
309 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
310 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
311 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
312 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
313 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
314 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
315 if XNN_LIKELY(nc >= 8) {
316 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
317 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
318 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
319 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
320 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
321 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
322
323 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
324 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
325 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
326 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
327 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
328 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
329
330 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
331 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
332 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
333 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
334 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
335 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
336
337 nc -= 8;
338 } else {
339 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
340 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
341 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
342 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
343 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
344 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
345 if (nc & 4) {
346 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
347 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
348 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
349 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
350 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
351 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
352 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
353 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
354 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
355 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
356 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
357 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
358 }
359 if (nc & 2) {
360 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
361 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
362 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
363 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
364 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
365 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
366 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
367 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
368 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
369 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
370 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
371 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
372 }
373 if (nc & 1) {
374 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
375 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
376 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
377 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
378 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
379 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
380 }
381 nc = 0;
382 }
383 } while (nc != 0);
384 }