Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-5x16c8-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_5x16c8__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 <= 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, 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 int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
72
73 // Loop over groups of 16 columns.
74 do {
75 // Initialize accumulators with scaled vksum. 16 scaled vksum 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_point45 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[4].zero_point));
100 float32x4_t vout4x0123 = vmulq_f32(vksum0123, vinput_zero_point45);
101 float32x4_t vout4x4567 = vmulq_f32(vksum4567, vinput_zero_point45);
102 float32x4_t vout4x89AB = vmulq_f32(vksum89AB, vinput_zero_point45);
103 float32x4_t vout4xCDEF = vmulq_f32(vksumCDEF, vinput_zero_point45);
104
105
106 for (size_t kb=0; kb < kc; kb += bl) {
107 int32x4_t vacc01x01 = vdupq_n_s32(0);
108 int32x4_t vacc01x23 = vdupq_n_s32(0);
109 int32x4_t vacc01x45 = vdupq_n_s32(0);
110 int32x4_t vacc01x67 = vdupq_n_s32(0);
111 int32x4_t vacc01x89 = vdupq_n_s32(0);
112 int32x4_t vacc01xAB = vdupq_n_s32(0);
113 int32x4_t vacc01xCD = vdupq_n_s32(0);
114 int32x4_t vacc01xEF = vdupq_n_s32(0);
115 int32x4_t vacc23x01 = vdupq_n_s32(0);
116 int32x4_t vacc23x23 = vdupq_n_s32(0);
117 int32x4_t vacc23x45 = vdupq_n_s32(0);
118 int32x4_t vacc23x67 = vdupq_n_s32(0);
119 int32x4_t vacc23x89 = vdupq_n_s32(0);
120 int32x4_t vacc23xAB = vdupq_n_s32(0);
121 int32x4_t vacc23xCD = vdupq_n_s32(0);
122 int32x4_t vacc23xEF = vdupq_n_s32(0);
123 int32x4_t vacc45x01 = vdupq_n_s32(0);
124 int32x4_t vacc45x23 = vdupq_n_s32(0);
125 int32x4_t vacc45x45 = vdupq_n_s32(0);
126 int32x4_t vacc45x67 = vdupq_n_s32(0);
127 int32x4_t vacc45x89 = vdupq_n_s32(0);
128 int32x4_t vacc45xAB = vdupq_n_s32(0);
129 int32x4_t vacc45xCD = vdupq_n_s32(0);
130 int32x4_t vacc45xEF = vdupq_n_s32(0);
131
132 size_t k = bl;
133 // 2x partial unrolled loop to load 8 bytes at a time.
134
135 uint64x2x2_t va01x0123456789ABCDEF;
136 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
137 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
138 uint64x2x2_t va23x0123456789ABCDEF;
139 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
140 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
141 uint64x2x2_t va45x0123456789ABCDEF;
142 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
143 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
144
145 while (k >= 16 * sizeof(int8_t)) {
146 // Load a 5x16 block of activations.
147 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
148 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
149 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
150 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
151 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
152
153 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
154 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
155 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
156 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
157 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
158 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
159
160 // Load a 16x16 block of weights.
161 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
162 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
163 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
164 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
165 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
166 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
167 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
168 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
169 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
170 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
171 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
172 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
173 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
174 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
175 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
176 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
177 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
178 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
179 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
180 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
181 const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask);
182 const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask);
183 const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask);
184 const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask);
185
186 // Multiply-accumulate: 5x8 * 8x16 --> 5x16.
187 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
188 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
189 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
190 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
191 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
192 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
193 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
194 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
195 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
196 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
197 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
198 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
199 vacc23x89 = vmmlaq_s32(vacc23x89, va23x01234567, vb89x01234567);
200 vacc23xAB = vmmlaq_s32(vacc23xAB, va23x01234567, vbABx01234567);
201 vacc23xCD = vmmlaq_s32(vacc23xCD, va23x01234567, vbCDx01234567);
202 vacc23xEF = vmmlaq_s32(vacc23xEF, va23x01234567, vbEFx01234567);
203 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
204 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
205 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
206 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
207 vacc45x89 = vmmlaq_s32(vacc45x89, va45x01234567, vb89x01234567);
208 vacc45xAB = vmmlaq_s32(vacc45xAB, va45x01234567, vbABx01234567);
209 vacc45xCD = vmmlaq_s32(vacc45xCD, va45x01234567, vbCDx01234567);
210 vacc45xEF = vmmlaq_s32(vacc45xEF, va45x01234567, vbEFx01234567);
211 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
212 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
213 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
214 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
215 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
216 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
217 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
218 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
219 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
220 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
221 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
222 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
223 vacc23x89 = vmmlaq_s32(vacc23x89, va23x89ABCDEF, vb89x89ABCDEF);
224 vacc23xAB = vmmlaq_s32(vacc23xAB, va23x89ABCDEF, vbABx89ABCDEF);
225 vacc23xCD = vmmlaq_s32(vacc23xCD, va23x89ABCDEF, vbCDx89ABCDEF);
226 vacc23xEF = vmmlaq_s32(vacc23xEF, va23x89ABCDEF, vbEFx89ABCDEF);
227 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
228 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
229 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
230 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
231 vacc45x89 = vmmlaq_s32(vacc45x89, va45x89ABCDEF, vb89x89ABCDEF);
232 vacc45xAB = vmmlaq_s32(vacc45xAB, va45x89ABCDEF, vbABx89ABCDEF);
233 vacc45xCD = vmmlaq_s32(vacc45xCD, va45x89ABCDEF, vbCDx89ABCDEF);
234 vacc45xEF = vmmlaq_s32(vacc45xEF, va45x89ABCDEF, vbEFx89ABCDEF);
235
236 k -= 16 * sizeof(int8_t);
237 }
238 // Handle up to 8 final positions of `k`
239 if XNN_UNLIKELY(k != 0) {
240 // Load a 5x8 block of activations.
241 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
242 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
243 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
244 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
245 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
246
247 // Load a 16x16 block of weights.
248 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
249 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
250 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
251 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
252 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
253 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
254 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
255 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
256 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
257 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
258 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
259 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
260 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
261 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
262 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
263 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
264
265 // Multiply-accumulate: 5x4 * 4x16 --> 5x16.
266 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
267 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
268 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
269 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
270 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
271 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
272 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
273 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
274 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
275 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
276 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
277 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
278 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x01234567), vb89x01234567);
279 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x01234567), vbABx01234567);
280 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x01234567), vbCDx01234567);
281 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x01234567), vbEFx01234567);
282 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
283 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
284 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
285 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
286 vacc45x89 = vmmlaq_s32(vacc45x89, vreinterpretq_s8_u64(va45x01234567), vb89x01234567);
287 vacc45xAB = vmmlaq_s32(vacc45xAB, vreinterpretq_s8_u64(va45x01234567), vbABx01234567);
288 vacc45xCD = vmmlaq_s32(vacc45xCD, vreinterpretq_s8_u64(va45x01234567), vbCDx01234567);
289 vacc45xEF = vmmlaq_s32(vacc45xEF, vreinterpretq_s8_u64(va45x01234567), vbEFx01234567);
290 }
291
292 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
293 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
294 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
295 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
296 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
297 int32x4_t vacc1x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
298 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
299 int32x4_t vacc1xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
300 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
301 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
302 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
303 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
304 int32x4_t vacc2x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x89), vreinterpretq_u64_s32(vacc23xAB)));
305 int32x4_t vacc3x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x89), vreinterpretq_u64_s32(vacc23xAB)));
306 int32x4_t vacc2xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23xCD), vreinterpretq_u64_s32(vacc23xEF)));
307 int32x4_t vacc3xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23xCD), vreinterpretq_u64_s32(vacc23xEF)));
308 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
309 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
310 int32x4_t vacc4x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x89), vreinterpretq_u64_s32(vacc45xAB)));
311 int32x4_t vacc4xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45xCD), vreinterpretq_u64_s32(vacc45xEF)));
312 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
313 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
314 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
315 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
316 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
317 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
318 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
319 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
320 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
321 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
322 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
323 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
324 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
325 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
326 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
327 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
328 float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
329 vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
330 float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
331 vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
332 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
333 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
334 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
335 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
336 float32x4_t vf2x89AB = vcvtq_f32_s32(vacc2x89AB);
337 vout2x89AB = vfmaq_f32(vout2x89AB, vf2x89AB, vfilter_output_scale89AB);
338 float32x4_t vf2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
339 vout2xCDEF = vfmaq_f32(vout2xCDEF, vf2xCDEF, vfilter_output_scaleCDEF);
340 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
341 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
342 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
343 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
344 float32x4_t vf3x89AB = vcvtq_f32_s32(vacc3x89AB);
345 vout3x89AB = vfmaq_f32(vout3x89AB, vf3x89AB, vfilter_output_scale89AB);
346 float32x4_t vf3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
347 vout3xCDEF = vfmaq_f32(vout3xCDEF, vf3xCDEF, vfilter_output_scaleCDEF);
348 float32x4_t vf4x0123 = vcvtq_f32_s32(vacc4x0123);
349 vout4x0123 = vfmaq_f32(vout4x0123, vf4x0123, vfilter_output_scale0123);
350 float32x4_t vf4x4567 = vcvtq_f32_s32(vacc4x4567);
351 vout4x4567 = vfmaq_f32(vout4x4567, vf4x4567, vfilter_output_scale4567);
352 float32x4_t vf4x89AB = vcvtq_f32_s32(vacc4x89AB);
353 vout4x89AB = vfmaq_f32(vout4x89AB, vf4x89AB, vfilter_output_scale89AB);
354 float32x4_t vf4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
355 vout4xCDEF = vfmaq_f32(vout4xCDEF, vf4xCDEF, vfilter_output_scaleCDEF);
356 }
357
358 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
359 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
360 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
361 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
362 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
363 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
364 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
365 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
366 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
367 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
368 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
369 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
370 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
371 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
372 vout2x89AB = vmulq_lane_f32(vout2x89AB, vget_low_f32(vinput_scale23), 1);
373 vout3x89AB = vmulq_lane_f32(vout3x89AB, vget_high_f32(vinput_scale23), 1);
374 vout2xCDEF = vmulq_lane_f32(vout2xCDEF, vget_low_f32(vinput_scale23), 1);
375 vout3xCDEF = vmulq_lane_f32(vout3xCDEF, vget_high_f32(vinput_scale23), 1);
376 const float32x4_t vinput_scale4 = vld1q_dup_f32(&quantization_params[4].inv_scale);
377 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale4);
378 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale4);
379 vout4x89AB = vmulq_f32(vout4x89AB, vinput_scale4);
380 vout4xCDEF = vmulq_f32(vout4xCDEF, vinput_scale4);
381
382
383 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
384 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
385 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
386 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
387 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
388 vout4x0123 = vaddq_f32(vbias0123, vout4x0123);
389 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
390 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
391 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
392 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
393 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
394 vout4x4567 = vaddq_f32(vbias4567, vout4x4567);
395 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
396 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
397 vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
398 vout2x89AB = vaddq_f32(vbias89AB, vout2x89AB);
399 vout3x89AB = vaddq_f32(vbias89AB, vout3x89AB);
400 vout4x89AB = vaddq_f32(vbias89AB, vout4x89AB);
401 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
402 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
403 vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);
404 vout2xCDEF = vaddq_f32(vbiasCDEF, vout2xCDEF);
405 vout3xCDEF = vaddq_f32(vbiasCDEF, vout3xCDEF);
406 vout4xCDEF = vaddq_f32(vbiasCDEF, vout4xCDEF);
407
408 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
409 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
410 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
411 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
412 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
413 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
414 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
415 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
416 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
417 float16x8_t vfp16out4x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout4x89AB), vcvt_f16_f32(vout4xCDEF));
418
419 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
420 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
421 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
422 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
423 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
424 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
425 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
426 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
427 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
428 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
429 vfp16out4x89ABCDEF = vmaxq_f16(vfp16out4x89ABCDEF, voutput_min);
430 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
431 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
432 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
433 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
434 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
435 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
436 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
437 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
438 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
439 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
440 vfp16out4x89ABCDEF = vminq_f16(vfp16out4x89ABCDEF, voutput_max);
441 if XNN_LIKELY(nc >= 16) {
442 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
443 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
444 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
445 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
446 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
447 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
448 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
449 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
450 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
451 vst1q_u16(c4 + 8, vreinterpretq_u16_f16(vfp16out4x89ABCDEF));
452
453 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
454 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
455 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
456 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
457 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
458
459 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
460 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
461 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
462 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
463 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
464
465 nc -= 16;
466 } else {
467 if (nc & 8) {
468 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
469 vfp16out0x01234567 = vfp16out0x89ABCDEF;
470 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
471 vfp16out1x01234567 = vfp16out1x89ABCDEF;
472 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
473 vfp16out2x01234567 = vfp16out2x89ABCDEF;
474 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
475 vfp16out3x01234567 = vfp16out3x89ABCDEF;
476 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567)); c4 += 8;
477 vfp16out4x01234567 = vfp16out4x89ABCDEF;
478 }
479 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
480 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
481 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
482 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
483 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
484 if (nc & 4) {
485 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
486 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
487 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
488 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
489 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
490 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
491 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
492 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
493 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
494 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
495 }
496 if (nc & 2) {
497 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
498 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
499 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
500 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
501 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
502 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
503 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
504 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
505 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
506 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
507 }
508 if (nc & 1) {
509 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
510 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
511 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
512 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
513 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
514 }
515 nc = 0;
516 }
517 } while (nc != 0);
518 }