Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-7x8c8-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_7x8c8__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 <= 7);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 size_t bl = params->scalar.blocksize;
42 assert(bl <= kc);
43 assert(bl != 0);
44 assert(bl % 32 == 0);
45 const int8_t* a0 = a;
46 uint16_t* c0 = (uint16_t*) c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr < 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
66 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 4) {
68 a4 = a3;
69 c4 = c3;
70 }
71 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
72 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 6) {
74 a5 = a4;
75 c5 = c4;
76 }
77 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
78 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
79 if XNN_UNPREDICTABLE(mr <= 6) {
80 a6 = a5;
81 c6 = c5;
82 }
83 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
84
85 // Loop over groups of 8 columns.
86 do {
87 // Initialize accumulators with scaled vksum. 8 scaled vksum values are loaded from the
88 // weight matrix, at the start of the group of 8 columns.
89 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
90 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
91 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
92 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
93 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
94 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
95 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, 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 const float32x4_t vinput_zero_point45 = vcvtq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
102 float32x4_t vout4x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point45), 0);
103 float32x4_t vout5x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point45), 0);
104 float32x4_t vout4x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point45), 0);
105 float32x4_t vout5x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point45), 0);
106 const float32x4_t vinput_zero_point67 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[6].zero_point));
107 float32x4_t vout6x0123 = vmulq_f32(vksum0123, vinput_zero_point67);
108 float32x4_t vout6x4567 = vmulq_f32(vksum4567, vinput_zero_point67);
109
110
111 for (size_t kb=0; kb < kc; kb += bl) {
112 int32x4_t vacc01x01 = vdupq_n_s32(0);
113 int32x4_t vacc01x23 = vdupq_n_s32(0);
114 int32x4_t vacc01x45 = vdupq_n_s32(0);
115 int32x4_t vacc01x67 = vdupq_n_s32(0);
116 int32x4_t vacc23x01 = vdupq_n_s32(0);
117 int32x4_t vacc23x23 = vdupq_n_s32(0);
118 int32x4_t vacc23x45 = vdupq_n_s32(0);
119 int32x4_t vacc23x67 = vdupq_n_s32(0);
120 int32x4_t vacc45x01 = vdupq_n_s32(0);
121 int32x4_t vacc45x23 = vdupq_n_s32(0);
122 int32x4_t vacc45x45 = vdupq_n_s32(0);
123 int32x4_t vacc45x67 = vdupq_n_s32(0);
124 int32x4_t vacc67x01 = vdupq_n_s32(0);
125 int32x4_t vacc67x23 = vdupq_n_s32(0);
126 int32x4_t vacc67x45 = vdupq_n_s32(0);
127 int32x4_t vacc67x67 = vdupq_n_s32(0);
128
129 size_t k = bl;
130 // 2x partial unrolled loop to load 8 bytes at a time.
131
132 uint64x2x2_t va01x0123456789ABCDEF;
133 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
134 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
135 uint64x2x2_t va23x0123456789ABCDEF;
136 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
137 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
138 uint64x2x2_t va45x0123456789ABCDEF;
139 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
140 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
141 uint64x2x2_t va67x0123456789ABCDEF;
142 va67x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
143 va67x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
144
145 while (k >= 16 * sizeof(int8_t)) {
146 // Load a 7x16 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 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a6, va67x0123456789ABCDEF, 0); a6 += 16;
151 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
152 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
153 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
154
155 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
156 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
157 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
158 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
159 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
160 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
161 const int8x16_t va67x01234567 = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[0]);
162 const int8x16_t va67x89ABCDEF = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[1]);
163
164 // Load a 16x8 block of weights.
165 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
166 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
167 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
168 const int8x16_t vb67x0123456789ABCDEF = 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 vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
174 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
175 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
176 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
177
178 // Multiply-accumulate: 7x8 * 8x8 --> 7x8.
179 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
180 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
181 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
182 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
183 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
184 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
185 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
186 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
187 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
188 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
189 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
190 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
191 vacc67x01 = vmmlaq_s32(vacc67x01, va67x01234567, vb01x01234567);
192 vacc67x23 = vmmlaq_s32(vacc67x23, va67x01234567, vb23x01234567);
193 vacc67x45 = vmmlaq_s32(vacc67x45, va67x01234567, vb45x01234567);
194 vacc67x67 = vmmlaq_s32(vacc67x67, va67x01234567, vb67x01234567);
195 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
196 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
197 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
198 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
199 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
200 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
201 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
202 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
203 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
204 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
205 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
206 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
207 vacc67x01 = vmmlaq_s32(vacc67x01, va67x89ABCDEF, vb01x89ABCDEF);
208 vacc67x23 = vmmlaq_s32(vacc67x23, va67x89ABCDEF, vb23x89ABCDEF);
209 vacc67x45 = vmmlaq_s32(vacc67x45, va67x89ABCDEF, vb45x89ABCDEF);
210 vacc67x67 = vmmlaq_s32(vacc67x67, va67x89ABCDEF, vb67x89ABCDEF);
211
212 k -= 16 * sizeof(int8_t);
213 }
214 // Handle up to 8 final positions of `k`
215 if XNN_UNLIKELY(k != 0) {
216 // Load a 7x8 block of activations.
217 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
218 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
219 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
220 uint64x2_t va67x01234567 = vld1q_dup_u64((const void*) a6); a6 += 8;
221 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
222 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
223 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
224
225 // Load a 16x8 block of weights.
226 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
227 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
228 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
229 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
230 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
231 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
232 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
233 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
234
235 // Multiply-accumulate: 7x4 * 4x8 --> 7x8.
236 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
237 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
238 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
239 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
240 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
241 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
242 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
243 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
244 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
245 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
246 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
247 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
248 vacc67x01 = vmmlaq_s32(vacc67x01, vreinterpretq_s8_u64(va67x01234567), vb01x01234567);
249 vacc67x23 = vmmlaq_s32(vacc67x23, vreinterpretq_s8_u64(va67x01234567), vb23x01234567);
250 vacc67x45 = vmmlaq_s32(vacc67x45, vreinterpretq_s8_u64(va67x01234567), vb45x01234567);
251 vacc67x67 = vmmlaq_s32(vacc67x67, vreinterpretq_s8_u64(va67x01234567), vb67x01234567);
252 }
253
254 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
255 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
256 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
257 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
258 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
259 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
260 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
261 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
262 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
263 int32x4_t vacc5x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
264 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
265 int32x4_t vacc5x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
266 int32x4_t vacc6x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
267 int32x4_t vacc6x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
268 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
269 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
270 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
271 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
272 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
273 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
274 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
275 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
276 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
277 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
278 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
279 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
280 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
281 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
282 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
283 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
284 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
285 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
286 float32x4_t vf4x0123 = vcvtq_f32_s32(vacc4x0123);
287 vout4x0123 = vfmaq_f32(vout4x0123, vf4x0123, vfilter_output_scale0123);
288 float32x4_t vf4x4567 = vcvtq_f32_s32(vacc4x4567);
289 vout4x4567 = vfmaq_f32(vout4x4567, vf4x4567, vfilter_output_scale4567);
290 float32x4_t vf5x0123 = vcvtq_f32_s32(vacc5x0123);
291 vout5x0123 = vfmaq_f32(vout5x0123, vf5x0123, vfilter_output_scale0123);
292 float32x4_t vf5x4567 = vcvtq_f32_s32(vacc5x4567);
293 vout5x4567 = vfmaq_f32(vout5x4567, vf5x4567, vfilter_output_scale4567);
294 float32x4_t vf6x0123 = vcvtq_f32_s32(vacc6x0123);
295 vout6x0123 = vfmaq_f32(vout6x0123, vf6x0123, vfilter_output_scale0123);
296 float32x4_t vf6x4567 = vcvtq_f32_s32(vacc6x4567);
297 vout6x4567 = vfmaq_f32(vout6x4567, vf6x4567, vfilter_output_scale4567);
298 }
299
300 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
301 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
302 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
303 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
304 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
305 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
306 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
307 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
308 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
309 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
310 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
311 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
312 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
313 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
314 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
315 const float32x4_t vinput_scale6 = vld1q_dup_f32(&quantization_params[6].inv_scale);
316 vout6x0123 = vmulq_f32(vout6x0123, vinput_scale6);
317 vout6x4567 = vmulq_f32(vout6x4567, vinput_scale6);
318
319
320 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
321 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
322 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
323 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
324 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
325 vout4x0123 = vaddq_f32(vbias0123, vout4x0123);
326 vout5x0123 = vaddq_f32(vbias0123, vout5x0123);
327 vout6x0123 = vaddq_f32(vbias0123, vout6x0123);
328 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
329 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
330 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
331 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
332 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
333 vout4x4567 = vaddq_f32(vbias4567, vout4x4567);
334 vout5x4567 = vaddq_f32(vbias4567, vout5x4567);
335 vout6x4567 = vaddq_f32(vbias4567, vout6x4567);
336
337 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
338 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
339 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
340 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
341 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
342 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
343 float16x8_t vfp16out6x01234567 = vcombine_f16(vcvt_f16_f32(vout6x0123), vcvt_f16_f32(vout6x4567));
344
345 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
346 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
347 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
348 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
349 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
350 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
351 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
352 vfp16out6x01234567 = vmaxq_f16(vfp16out6x01234567, voutput_min);
353 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
354 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
355 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
356 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
357 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
358 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
359 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
360 vfp16out6x01234567 = vminq_f16(vfp16out6x01234567, voutput_max);
361 if XNN_LIKELY(nc >= 8) {
362 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
363 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
364 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
365 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
366 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
367 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
368 vst1q_u16(c6, vreinterpretq_u16_f16(vfp16out6x01234567));
369
370 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
371 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
372 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
373 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
374 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
375 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
376 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
377
378 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
379 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
380 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
381 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
382 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
383 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
384 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
385
386 nc -= 8;
387 } else {
388 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
389 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
390 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
391 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
392 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
393 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
394 float16x4_t vfp16out6x0123 = vget_low_f16(vfp16out6x01234567);
395 if (nc & 4) {
396 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
397 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
398 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
399 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
400 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
401 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
402 vst1_u16(c6, vreinterpret_u16_f16(vfp16out6x0123)); c6 += 4;
403 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
404 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
405 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
406 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
407 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
408 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
409 vfp16out6x0123 = vget_high_f16(vfp16out6x01234567);
410 }
411 if (nc & 2) {
412 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
413 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
414 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
415 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
416 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
417 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
418 vst1_lane_u32((void*) c6, vreinterpret_u32_f16(vfp16out6x0123), 0); c6 += 2;
419 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
420 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
421 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
422 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
423 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
424 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
425 vfp16out6x0123 = vext_f16(vfp16out6x0123, vfp16out6x0123, 2);
426 }
427 if (nc & 1) {
428 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
429 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
430 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
431 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
432 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
433 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
434 vst1_lane_u16(c6, vreinterpret_u16_f16(vfp16out6x0123), 0);
435 }
436 nc = 0;
437 }
438 } while (nc != 0);
439 }