Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-5x16c4-minmax-neondot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-neondot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_qb4w_gemm_minmax_ukernel_5x16c4__neondot(
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 struct xnn_f32_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 <= 5);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 4 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67
68 size_t bl = params->scalar.blocksize;
69 assert(bl <= kc);
70 assert(bl != 0);
71 assert(bl % 32 == 0);
72 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
73 // Loop over groups of 16 columns.
74 do {
75 // Initialize accumulators with bias. 16 bias values are loaded from the
76 // weight matrix, at the start of the group of 16 columns.
77 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
78 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
79 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
80 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
81 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
82 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
83 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
84 const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
85 float32x4_t vout0x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point01), 0);
86 float32x4_t vout1x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point01), 0);
87 const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
88 float32x4_t vout0xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point01), 0);
89 float32x4_t vout1xCDEF = vmulq_lane_f32(vksumCDEF, 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 float32x4_t vout2x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point23), 0);
96 float32x4_t vout3x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point23), 0);
97 float32x4_t vout2xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point23), 0);
98 float32x4_t vout3xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point23), 0);
99 const float32x4_t vinput_zero_point4 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[4].zero_point));
100 float32x4_t vout4x0123 = vmulq_f32(vksum0123, vinput_zero_point4);
101 float32x4_t vout4x4567 = vmulq_f32(vksum4567, vinput_zero_point4);
102 float32x4_t vout4x89AB = vmulq_f32(vksum89AB, vinput_zero_point4);
103 float32x4_t vout4xCDEF = vmulq_f32(vksumCDEF, vinput_zero_point4);
104
105 for (size_t kb=0; kb < kc; kb += bl) {
106 int32x4_t vacc0x0123 = vdupq_n_s32(0);
107 int32x4_t vacc1x0123 = vdupq_n_s32(0);
108 int32x4_t vacc0x4567 = vdupq_n_s32(0);
109 int32x4_t vacc1x4567 = vdupq_n_s32(0);
110 int32x4_t vacc0x89AB = vdupq_n_s32(0);
111 int32x4_t vacc1x89AB = vdupq_n_s32(0);
112 int32x4_t vacc0xCDEF = vdupq_n_s32(0);
113 int32x4_t vacc1xCDEF = vdupq_n_s32(0);
114 int32x4_t vacc2x0123 = vdupq_n_s32(0);
115 int32x4_t vacc3x0123 = vdupq_n_s32(0);
116 int32x4_t vacc2x4567 = vdupq_n_s32(0);
117 int32x4_t vacc3x4567 = vdupq_n_s32(0);
118 int32x4_t vacc2x89AB = vdupq_n_s32(0);
119 int32x4_t vacc3x89AB = vdupq_n_s32(0);
120 int32x4_t vacc2xCDEF = vdupq_n_s32(0);
121 int32x4_t vacc3xCDEF = vdupq_n_s32(0);
122 int32x4_t vacc4x0123 = vdupq_n_s32(0);
123 int32x4_t vacc4x4567 = vdupq_n_s32(0);
124 int32x4_t vacc4x89AB = vdupq_n_s32(0);
125 int32x4_t vacc4xCDEF = vdupq_n_s32(0);
126 // Inner accumulation loop along the 16 columns.
127 size_t k = bl;
128 // 2x partial unrolled loop to load 8 bytes at a time.
129 while (k >= 8 * sizeof(int8_t)) {
130 // Load a 5x8 block of activations.
131 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
132 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
133 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
134 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
135 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
136
137 // Load a 8x16 block of weights.
138 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
139 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
140 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
141 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
142 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
143 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
144 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
145 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
146 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
147 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
148 const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
149 const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);
150
151 // Multiply-accumulate: 5x8 * 8x16 --> 5x16.
152 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
153 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
154 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
155 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
156 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
157 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
158 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
159 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
160 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
161 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
162 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
163 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
164 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
165 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
166 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
167 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
168 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
169 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
170 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
171 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
172 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
173 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
174 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
175 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
176 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
177 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
178 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
179 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
180 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
181 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
182 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
183 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
184 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
185 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
186 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
187 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
188 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
189 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
190 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
191 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
192
193 k -= 8 * sizeof(int8_t);
194 }
195 // Handle up to 4 final positions of `k`
196 if XNN_UNLIKELY(k != 0) {
197 // Load a 5x4 block of activations.
198 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
199 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
200 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
201 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
202 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
203
204 // Load a 4x16 block of weights.
205 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
206 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
207 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
208 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
209
210 // Multiply-accumulate: 5x4 * 4x16 --> 5x16.
211 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
212 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
213 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
214 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
215 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
216 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
217 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
218 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
219 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
220 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
221 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
222 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
223 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
224 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
225 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
226 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
227 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
228 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
229 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
230 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
231 }
232 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
233 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
234 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
235 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
236
237 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
238 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
239 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
240 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
241 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
242 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
243 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
244 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
245 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
246 float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
247 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
248 vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
249 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
250 float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
251 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
252 vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
253 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
254 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
255 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
256 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
257 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
258 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
259 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
260 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
261 float32x4_t vf2x89AB = vcvtq_f32_s32(vacc2x89AB);
262 float32x4_t vf3x89AB = vcvtq_f32_s32(vacc3x89AB);
263 vout2x89AB = vfmaq_f32(vout2x89AB, vf2x89AB, vfilter_output_scale89AB);
264 vout3x89AB = vfmaq_f32(vout3x89AB, vf3x89AB, vfilter_output_scale89AB);
265 float32x4_t vf2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
266 float32x4_t vf3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
267 vout2xCDEF = vfmaq_f32(vout2xCDEF, vf2xCDEF, vfilter_output_scaleCDEF);
268 vout3xCDEF = vfmaq_f32(vout3xCDEF, vf3xCDEF, vfilter_output_scaleCDEF);
269 float32x4_t vf4x0123 = vcvtq_f32_s32(vacc4x0123);
270 vout4x0123 = vfmaq_f32(vout4x0123, vf4x0123, vfilter_output_scale0123);
271 float32x4_t vf4x4567 = vcvtq_f32_s32(vacc4x4567);
272 vout4x4567 = vfmaq_f32(vout4x4567, vf4x4567, vfilter_output_scale4567);
273 float32x4_t vf4x89AB = vcvtq_f32_s32(vacc4x89AB);
274 vout4x89AB = vfmaq_f32(vout4x89AB, vf4x89AB, vfilter_output_scale89AB);
275 float32x4_t vf4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
276 vout4xCDEF = vfmaq_f32(vout4xCDEF, vf4xCDEF, vfilter_output_scaleCDEF);
277 }
278
279 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
280 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
281 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
282 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
283 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
284 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
285 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
286 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
287 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
288 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
289 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
290 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
291 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
292 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
293 vout2x89AB = vmulq_lane_f32(vout2x89AB, vget_low_f32(vinput_scale23), 1);
294 vout3x89AB = vmulq_lane_f32(vout3x89AB, vget_high_f32(vinput_scale23), 1);
295 vout2xCDEF = vmulq_lane_f32(vout2xCDEF, vget_low_f32(vinput_scale23), 1);
296 vout3xCDEF = vmulq_lane_f32(vout3xCDEF, vget_high_f32(vinput_scale23), 1);
297 const float32x4_t vinput_scale4 = vld1q_dup_f32(&quantization_params[4].inv_scale);
298 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale4);
299 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale4);
300 vout4x89AB = vmulq_f32(vout4x89AB, vinput_scale4);
301 vout4xCDEF = vmulq_f32(vout4xCDEF, vinput_scale4);
302
303
304 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
305 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
306 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
307 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
308 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
309 vout4x0123 = vaddq_f32(vbias0123, vout4x0123);
310 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
311 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
312 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
313 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
314 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
315 vout4x4567 = vaddq_f32(vbias4567, vout4x4567);
316 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
317 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
318 vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
319 vout2x89AB = vaddq_f32(vbias89AB, vout2x89AB);
320 vout3x89AB = vaddq_f32(vbias89AB, vout3x89AB);
321 vout4x89AB = vaddq_f32(vbias89AB, vout4x89AB);
322 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
323 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
324 vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);
325 vout2xCDEF = vaddq_f32(vbiasCDEF, vout2xCDEF);
326 vout3xCDEF = vaddq_f32(vbiasCDEF, vout3xCDEF);
327 vout4xCDEF = vaddq_f32(vbiasCDEF, vout4xCDEF);
328
329 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
330 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
331 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
332 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
333 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
334 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
335 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
336 vout1x89AB = vmaxq_f32(vout1x89AB, voutput_min);
337 vout1xCDEF = vmaxq_f32(vout1xCDEF, voutput_min);
338 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
339 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
340 vout2x89AB = vmaxq_f32(vout2x89AB, voutput_min);
341 vout2xCDEF = vmaxq_f32(vout2xCDEF, voutput_min);
342 vout3x0123 = vmaxq_f32(vout3x0123, voutput_min);
343 vout3x4567 = vmaxq_f32(vout3x4567, voutput_min);
344 vout3x89AB = vmaxq_f32(vout3x89AB, voutput_min);
345 vout3xCDEF = vmaxq_f32(vout3xCDEF, voutput_min);
346 vout4x0123 = vmaxq_f32(vout4x0123, voutput_min);
347 vout4x4567 = vmaxq_f32(vout4x4567, voutput_min);
348 vout4x89AB = vmaxq_f32(vout4x89AB, voutput_min);
349 vout4xCDEF = vmaxq_f32(vout4xCDEF, voutput_min);
350
351 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
352 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
353 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
354 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
355 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
356 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
357 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
358 vout1x89AB = vminq_f32(vout1x89AB, voutput_max);
359 vout1xCDEF = vminq_f32(vout1xCDEF, voutput_max);
360 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
361 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
362 vout2x89AB = vminq_f32(vout2x89AB, voutput_max);
363 vout2xCDEF = vminq_f32(vout2xCDEF, voutput_max);
364 vout3x0123 = vminq_f32(vout3x0123, voutput_max);
365 vout3x4567 = vminq_f32(vout3x4567, voutput_max);
366 vout3x89AB = vminq_f32(vout3x89AB, voutput_max);
367 vout3xCDEF = vminq_f32(vout3xCDEF, voutput_max);
368 vout4x0123 = vminq_f32(vout4x0123, voutput_max);
369 vout4x4567 = vminq_f32(vout4x4567, voutput_max);
370 vout4x89AB = vminq_f32(vout4x89AB, voutput_max);
371 vout4xCDEF = vminq_f32(vout4xCDEF, voutput_max);
372
373 if XNN_LIKELY(nc >= 16) {
374 vst1q_f32(c0, vout0x0123);
375 vst1q_f32(c0 + 4, vout0x4567);
376 vst1q_f32(c0 + 8, vout0x89AB);
377 vst1q_f32(c0 + 12, vout0xCDEF);
378 vst1q_f32(c1, vout1x0123);
379 vst1q_f32(c1 + 4, vout1x4567);
380 vst1q_f32(c1 + 8, vout1x89AB);
381 vst1q_f32(c1 + 12, vout1xCDEF);
382 vst1q_f32(c2, vout2x0123);
383 vst1q_f32(c2 + 4, vout2x4567);
384 vst1q_f32(c2 + 8, vout2x89AB);
385 vst1q_f32(c2 + 12, vout2xCDEF);
386 vst1q_f32(c3, vout3x0123);
387 vst1q_f32(c3 + 4, vout3x4567);
388 vst1q_f32(c3 + 8, vout3x89AB);
389 vst1q_f32(c3 + 12, vout3xCDEF);
390 vst1q_f32(c4, vout4x0123);
391 vst1q_f32(c4 + 4, vout4x4567);
392 vst1q_f32(c4 + 8, vout4x89AB);
393 vst1q_f32(c4 + 12, vout4xCDEF);
394
395 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
396 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
397 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
398 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
399 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
400
401 c0 = (float*) ((uintptr_t) c0 + cn_stride);
402 c1 = (float*) ((uintptr_t) c1 + cn_stride);
403 c2 = (float*) ((uintptr_t) c2 + cn_stride);
404 c3 = (float*) ((uintptr_t) c3 + cn_stride);
405 c4 = (float*) ((uintptr_t) c4 + cn_stride);
406
407 nc -= 16;
408 } else {
409 if (nc & 8) {
410 vst1q_f32(c0, vout0x0123); c0 += 4;
411 vout0x0123 = vout0x89AB;
412 vst1q_f32(c1, vout1x0123); c1 += 4;
413 vout1x0123 = vout1x89AB;
414 vst1q_f32(c2, vout2x0123); c2 += 4;
415 vout2x0123 = vout2x89AB;
416 vst1q_f32(c3, vout3x0123); c3 += 4;
417 vout3x0123 = vout3x89AB;
418 vst1q_f32(c4, vout4x0123); c4 += 4;
419 vout4x0123 = vout4x89AB;
420 vst1q_f32(c0, vout0x4567); c0 += 4;
421 vout0x4567 = vout0xCDEF;
422 vst1q_f32(c1, vout1x4567); c1 += 4;
423 vout1x4567 = vout1xCDEF;
424 vst1q_f32(c2, vout2x4567); c2 += 4;
425 vout2x4567 = vout2xCDEF;
426 vst1q_f32(c3, vout3x4567); c3 += 4;
427 vout3x4567 = vout3xCDEF;
428 vst1q_f32(c4, vout4x4567); c4 += 4;
429 vout4x4567 = vout4xCDEF;
430 }
431 if (nc & 4) {
432 vst1q_f32(c0, vout0x0123); c0 += 4;
433 vout0x0123 = vout0x4567;
434 vst1q_f32(c1, vout1x0123); c1 += 4;
435 vout1x0123 = vout1x4567;
436 vst1q_f32(c2, vout2x0123); c2 += 4;
437 vout2x0123 = vout2x4567;
438 vst1q_f32(c3, vout3x0123); c3 += 4;
439 vout3x0123 = vout3x4567;
440 vst1q_f32(c4, vout4x0123); c4 += 4;
441 vout4x0123 = vout4x4567;
442 }
443 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
444 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
445 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
446 float32x2_t vout3x01 = vget_low_f32(vout3x0123);
447 float32x2_t vout4x01 = vget_low_f32(vout4x0123);
448 if (nc & 2) {
449 vst1_f32(c0, vout0x01); c0 += 2;
450 vst1_f32(c1, vout1x01); c1 += 2;
451 vst1_f32(c2, vout2x01); c2 += 2;
452 vst1_f32(c3, vout3x01); c3 += 2;
453 vst1_f32(c4, vout4x01); c4 += 2;
454 vout0x01 = vget_high_f32(vout0x0123);
455 vout1x01 = vget_high_f32(vout1x0123);
456 vout2x01 = vget_high_f32(vout2x0123);
457 vout3x01 = vget_high_f32(vout3x0123);
458 vout4x01 = vget_high_f32(vout4x0123);
459 }
460 if (nc & 1) {
461 vst1_lane_f32(c0, vout0x01, 0);
462 vst1_lane_f32(c1, vout1x01, 0);
463 vst1_lane_f32(c2, vout2x01, 0);
464 vst1_lane_f32(c3, vout3x01, 0);
465 vst1_lane_f32(c4, vout4x01, 0);
466 }
467 nc = 0;
468 }
469 } while (nc != 0);
470 }