Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-6x16c4-minmax-neondotfp16arith.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_f16_qb4w_gemm_minmax_ukernel_6x16c4__neondotfp16arith(
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, 4 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 uint16_t* c0 = (uint16_t*) c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 uint16_t* c1 = (uint16_t*) ((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 uint16_t* c2 = (uint16_t*) ((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 uint16_t* c3 = (uint16_t*) ((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 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr != 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73
74 size_t bl = params->scalar.blocksize;
75 assert(bl <= kc);
76 assert(bl != 0);
77 assert(bl % 32 == 0);
78 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
79 // Loop over groups of 16 columns.
80 do {
81 // Initialize accumulators with bias. 16 bias values are loaded from the
82 // weight matrix, at the start of the group of 16 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 vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
91 float32x4_t vout0x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point01), 0);
92 float32x4_t vout1x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point01), 0);
93 const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
94 float32x4_t vout0xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point01), 0);
95 float32x4_t vout1xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point01), 0);
96 const float32x4_t vinput_zero_point23 = vcvtq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
97 float32x4_t vout2x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point23), 0);
98 float32x4_t vout3x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point23), 0);
99 float32x4_t vout2x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point23), 0);
100 float32x4_t vout3x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point23), 0);
101 float32x4_t vout2x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point23), 0);
102 float32x4_t vout3x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point23), 0);
103 float32x4_t vout2xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point23), 0);
104 float32x4_t vout3xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point23), 0);
105 const float32x4_t vinput_zero_point45 = vcvtq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
106 float32x4_t vout4x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point45), 0);
107 float32x4_t vout5x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point45), 0);
108 float32x4_t vout4x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point45), 0);
109 float32x4_t vout5x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point45), 0);
110 float32x4_t vout4x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point45), 0);
111 float32x4_t vout5x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point45), 0);
112 float32x4_t vout4xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point45), 0);
113 float32x4_t vout5xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point45), 0);
114
115 for (size_t kb=0; kb < kc; kb += bl) {
116 int32x4_t vacc0x0123 = vdupq_n_s32(0);
117 int32x4_t vacc1x0123 = vdupq_n_s32(0);
118 int32x4_t vacc0x4567 = vdupq_n_s32(0);
119 int32x4_t vacc1x4567 = vdupq_n_s32(0);
120 int32x4_t vacc0x89AB = vdupq_n_s32(0);
121 int32x4_t vacc1x89AB = vdupq_n_s32(0);
122 int32x4_t vacc0xCDEF = vdupq_n_s32(0);
123 int32x4_t vacc1xCDEF = vdupq_n_s32(0);
124 int32x4_t vacc2x0123 = vdupq_n_s32(0);
125 int32x4_t vacc3x0123 = vdupq_n_s32(0);
126 int32x4_t vacc2x4567 = vdupq_n_s32(0);
127 int32x4_t vacc3x4567 = vdupq_n_s32(0);
128 int32x4_t vacc2x89AB = vdupq_n_s32(0);
129 int32x4_t vacc3x89AB = vdupq_n_s32(0);
130 int32x4_t vacc2xCDEF = vdupq_n_s32(0);
131 int32x4_t vacc3xCDEF = vdupq_n_s32(0);
132 int32x4_t vacc4x0123 = vdupq_n_s32(0);
133 int32x4_t vacc5x0123 = vdupq_n_s32(0);
134 int32x4_t vacc4x4567 = vdupq_n_s32(0);
135 int32x4_t vacc5x4567 = vdupq_n_s32(0);
136 int32x4_t vacc4x89AB = vdupq_n_s32(0);
137 int32x4_t vacc5x89AB = vdupq_n_s32(0);
138 int32x4_t vacc4xCDEF = vdupq_n_s32(0);
139 int32x4_t vacc5xCDEF = vdupq_n_s32(0);
140 // Inner accumulation loop along the 16 columns.
141 size_t k = bl;
142 // 2x partial unrolled loop to load 8 bytes at a time.
143 while (k >= 8 * sizeof(int8_t)) {
144 // Load a 6x8 block of activations.
145 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
146 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
147 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
148 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
149 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
150 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
151
152 // Load a 8x16 block of weights.
153 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
154 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
155 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
156 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
157 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
158 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
159 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
160 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
161 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
162 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
163 const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
164 const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);
165
166 // Multiply-accumulate: 6x8 * 8x16 --> 6x16.
167 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
168 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
169 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
170 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
171 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
172 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
173 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
174 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
175 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
176 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
177 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
178 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
179 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
180 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
181 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
182 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
183 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
184 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
185 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
186 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
187 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
188 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
189 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
190 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
191 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
192 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
193 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
194 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
195 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
196 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
197 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
198 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
199 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
200 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
201 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
202 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
203 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
204 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
205 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
206 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
207 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
208 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
209 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
210 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
211 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
212 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
213 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1);
214 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1);
215
216 k -= 8 * sizeof(int8_t);
217 }
218 // Handle up to 4 final positions of `k`
219 if XNN_UNLIKELY(k != 0) {
220 // Load a 6x4 block of activations.
221 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
222 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
223 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
224 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
225 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
226 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4;
227
228 // Load a 4x16 block of weights.
229 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
230 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
231 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
232 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
233
234 // Multiply-accumulate: 6x4 * 4x16 --> 6x16.
235 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
236 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
237 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
238 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
239 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
240 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
241 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
242 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
243 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
244 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
245 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
246 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
247 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
248 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
249 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
250 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
251 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
252 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
253 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
254 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
255 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
256 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
257 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
258 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
259 }
260 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
261 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
262 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
263 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
264
265 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
266 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
267 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
268 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
269 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
270 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
271 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
272 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
273 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
274 float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
275 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
276 vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
277 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
278 float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
279 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
280 vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
281 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
282 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
283 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
284 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
285 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
286 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
287 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
288 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
289 float32x4_t vf2x89AB = vcvtq_f32_s32(vacc2x89AB);
290 float32x4_t vf3x89AB = vcvtq_f32_s32(vacc3x89AB);
291 vout2x89AB = vfmaq_f32(vout2x89AB, vf2x89AB, vfilter_output_scale89AB);
292 vout3x89AB = vfmaq_f32(vout3x89AB, vf3x89AB, vfilter_output_scale89AB);
293 float32x4_t vf2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
294 float32x4_t vf3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
295 vout2xCDEF = vfmaq_f32(vout2xCDEF, vf2xCDEF, vfilter_output_scaleCDEF);
296 vout3xCDEF = vfmaq_f32(vout3xCDEF, vf3xCDEF, vfilter_output_scaleCDEF);
297 float32x4_t vf4x0123 = vcvtq_f32_s32(vacc4x0123);
298 float32x4_t vf5x0123 = vcvtq_f32_s32(vacc5x0123);
299 vout4x0123 = vfmaq_f32(vout4x0123, vf4x0123, vfilter_output_scale0123);
300 vout5x0123 = vfmaq_f32(vout5x0123, vf5x0123, vfilter_output_scale0123);
301 float32x4_t vf4x4567 = vcvtq_f32_s32(vacc4x4567);
302 float32x4_t vf5x4567 = vcvtq_f32_s32(vacc5x4567);
303 vout4x4567 = vfmaq_f32(vout4x4567, vf4x4567, vfilter_output_scale4567);
304 vout5x4567 = vfmaq_f32(vout5x4567, vf5x4567, vfilter_output_scale4567);
305 float32x4_t vf4x89AB = vcvtq_f32_s32(vacc4x89AB);
306 float32x4_t vf5x89AB = vcvtq_f32_s32(vacc5x89AB);
307 vout4x89AB = vfmaq_f32(vout4x89AB, vf4x89AB, vfilter_output_scale89AB);
308 vout5x89AB = vfmaq_f32(vout5x89AB, vf5x89AB, vfilter_output_scale89AB);
309 float32x4_t vf4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
310 float32x4_t vf5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
311 vout4xCDEF = vfmaq_f32(vout4xCDEF, vf4xCDEF, vfilter_output_scaleCDEF);
312 vout5xCDEF = vfmaq_f32(vout5xCDEF, vf5xCDEF, vfilter_output_scaleCDEF);
313 }
314
315 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
316 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
317 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
318 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
319 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
320 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
321 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
322 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
323 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
324 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
325 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
326 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
327 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
328 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
329 vout2x89AB = vmulq_lane_f32(vout2x89AB, vget_low_f32(vinput_scale23), 1);
330 vout3x89AB = vmulq_lane_f32(vout3x89AB, vget_high_f32(vinput_scale23), 1);
331 vout2xCDEF = vmulq_lane_f32(vout2xCDEF, vget_low_f32(vinput_scale23), 1);
332 vout3xCDEF = vmulq_lane_f32(vout3xCDEF, vget_high_f32(vinput_scale23), 1);
333 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
334 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
335 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
336 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
337 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
338 vout4x89AB = vmulq_lane_f32(vout4x89AB, vget_low_f32(vinput_scale45), 1);
339 vout5x89AB = vmulq_lane_f32(vout5x89AB, vget_high_f32(vinput_scale45), 1);
340 vout4xCDEF = vmulq_lane_f32(vout4xCDEF, vget_low_f32(vinput_scale45), 1);
341 vout5xCDEF = vmulq_lane_f32(vout5xCDEF, vget_high_f32(vinput_scale45), 1);
342
343
344 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
345 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
346 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
347 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
348 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
349 vout4x0123 = vaddq_f32(vbias0123, vout4x0123);
350 vout5x0123 = vaddq_f32(vbias0123, vout5x0123);
351 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
352 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
353 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
354 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
355 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
356 vout4x4567 = vaddq_f32(vbias4567, vout4x4567);
357 vout5x4567 = vaddq_f32(vbias4567, vout5x4567);
358 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
359 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
360 vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
361 vout2x89AB = vaddq_f32(vbias89AB, vout2x89AB);
362 vout3x89AB = vaddq_f32(vbias89AB, vout3x89AB);
363 vout4x89AB = vaddq_f32(vbias89AB, vout4x89AB);
364 vout5x89AB = vaddq_f32(vbias89AB, vout5x89AB);
365 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
366 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
367 vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);
368 vout2xCDEF = vaddq_f32(vbiasCDEF, vout2xCDEF);
369 vout3xCDEF = vaddq_f32(vbiasCDEF, vout3xCDEF);
370 vout4xCDEF = vaddq_f32(vbiasCDEF, vout4xCDEF);
371 vout5xCDEF = vaddq_f32(vbiasCDEF, vout5xCDEF);
372
373 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
374 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
375 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
376 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
377 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
378 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
379 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
380 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
381 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
382 float16x8_t vfp16out4x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout4x89AB), vcvt_f16_f32(vout4xCDEF));
383 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
384 float16x8_t vfp16out5x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout5x89AB), vcvt_f16_f32(vout5xCDEF));
385
386 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
387 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
388 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
389 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
390 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
391 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
392 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
393 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
394 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
395 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
396 vfp16out4x89ABCDEF = vmaxq_f16(vfp16out4x89ABCDEF, voutput_min);
397 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
398 vfp16out5x89ABCDEF = vmaxq_f16(vfp16out5x89ABCDEF, voutput_min);
399 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
400 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
401 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
402 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
403 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
404 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
405 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
406 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
407 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
408 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
409 vfp16out4x89ABCDEF = vminq_f16(vfp16out4x89ABCDEF, voutput_max);
410 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
411 vfp16out5x89ABCDEF = vminq_f16(vfp16out5x89ABCDEF, voutput_max);
412 if XNN_LIKELY(nc >= 16) {
413 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
414 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
415 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
416 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
417 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
418 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
419 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
420 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
421 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
422 vst1q_u16(c4 + 8, vreinterpretq_u16_f16(vfp16out4x89ABCDEF));
423 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
424 vst1q_u16(c5 + 8, vreinterpretq_u16_f16(vfp16out5x89ABCDEF));
425
426 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
427 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
428 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
429 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
430 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
431 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
432
433 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
434 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
435 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
436 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
437 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
438 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
439
440 nc -= 16;
441 } else {
442 if (nc & 8) {
443 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
444 vfp16out0x01234567 = vfp16out0x89ABCDEF;
445 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
446 vfp16out1x01234567 = vfp16out1x89ABCDEF;
447 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
448 vfp16out2x01234567 = vfp16out2x89ABCDEF;
449 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
450 vfp16out3x01234567 = vfp16out3x89ABCDEF;
451 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567)); c4 += 8;
452 vfp16out4x01234567 = vfp16out4x89ABCDEF;
453 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567)); c5 += 8;
454 vfp16out5x01234567 = vfp16out5x89ABCDEF;
455 }
456 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
457 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
458 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
459 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
460 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
461 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
462 if (nc & 4) {
463 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
464 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
465 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
466 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
467 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
468 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
469 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
470 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
471 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
472 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
473 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
474 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
475 }
476 if (nc & 2) {
477 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
478 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
479 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
480 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
481 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
482 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
483 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
484 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
485 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
486 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
487 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
488 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
489 }
490 if (nc & 1) {
491 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
492 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
493 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
494 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
495 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
496 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
497 }
498 nc = 0;
499 }
500 } while (nc != 0);
501 }