Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-2x32c8-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_2x32c8__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 <= 2);
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 int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
54
55 // Loop over groups of 32 columns.
56 do {
57 // Initialize accumulators with scaled vksum. 32 scaled vksum values are loaded from the
58 // weight matrix, at the start of the group of 32 columns.
59 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
60 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
61 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
62 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
63 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
64 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
65 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
66 const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
67 float32x4_t vout0x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point01), 0);
68 float32x4_t vout1x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point01), 0);
69 const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
70 float32x4_t vout0xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point01), 0);
71 float32x4_t vout1xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point01), 0);
72 const float32x4_t vksumGHIJ = vld1q_f32(w); w = (const float*) w + 4;
73 float32x4_t vout0xGHIJ = vmulq_lane_f32(vksumGHIJ, vget_low_f32(vinput_zero_point01), 0);
74 float32x4_t vout1xGHIJ = vmulq_lane_f32(vksumGHIJ, vget_high_f32(vinput_zero_point01), 0);
75 const float32x4_t vksumKLMN = vld1q_f32(w); w = (const float*) w + 4;
76 float32x4_t vout0xKLMN = vmulq_lane_f32(vksumKLMN, vget_low_f32(vinput_zero_point01), 0);
77 float32x4_t vout1xKLMN = vmulq_lane_f32(vksumKLMN, vget_high_f32(vinput_zero_point01), 0);
78 const float32x4_t vksumOPQR = vld1q_f32(w); w = (const float*) w + 4;
79 float32x4_t vout0xOPQR = vmulq_lane_f32(vksumOPQR, vget_low_f32(vinput_zero_point01), 0);
80 float32x4_t vout1xOPQR = vmulq_lane_f32(vksumOPQR, vget_high_f32(vinput_zero_point01), 0);
81 const float32x4_t vksumSTUV = vld1q_f32(w); w = (const float*) w + 4;
82 float32x4_t vout0xSTUV = vmulq_lane_f32(vksumSTUV, vget_low_f32(vinput_zero_point01), 0);
83 float32x4_t vout1xSTUV = vmulq_lane_f32(vksumSTUV, vget_high_f32(vinput_zero_point01), 0);
84
85
86 for (size_t kb=0; kb < kc; kb += bl) {
87 int32x4_t vacc01x01 = vdupq_n_s32(0);
88 int32x4_t vacc01x23 = vdupq_n_s32(0);
89 int32x4_t vacc01x45 = vdupq_n_s32(0);
90 int32x4_t vacc01x67 = vdupq_n_s32(0);
91 int32x4_t vacc01x89 = vdupq_n_s32(0);
92 int32x4_t vacc01xAB = vdupq_n_s32(0);
93 int32x4_t vacc01xCD = vdupq_n_s32(0);
94 int32x4_t vacc01xEF = vdupq_n_s32(0);
95 int32x4_t vacc01xGH = vdupq_n_s32(0);
96 int32x4_t vacc01xIJ = vdupq_n_s32(0);
97 int32x4_t vacc01xKL = vdupq_n_s32(0);
98 int32x4_t vacc01xMN = vdupq_n_s32(0);
99 int32x4_t vacc01xOP = vdupq_n_s32(0);
100 int32x4_t vacc01xQR = vdupq_n_s32(0);
101 int32x4_t vacc01xST = vdupq_n_s32(0);
102 int32x4_t vacc01xUV = vdupq_n_s32(0);
103
104 size_t k = bl;
105 // 2x partial unrolled loop to load 8 bytes at a time.
106
107 uint64x2x2_t va01x0123456789ABCDEF;
108 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
109 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
110
111 while (k >= 16 * sizeof(int8_t)) {
112 // Load a 2x16 block of activations.
113 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
114 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
115
116 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
117 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
118
119 // Load a 16x32 block of weights.
120 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
121 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
122 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
123 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
124 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
125 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
126 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
127 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
128 const int8x16_t vbGHx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
129 const int8x16_t vbIJx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
130 const int8x16_t vbKLx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
131 const int8x16_t vbMNx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
132 const int8x16_t vbOPx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
133 const int8x16_t vbQRx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
134 const int8x16_t vbSTx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
135 const int8x16_t vbUVx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
136 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
137 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
138 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
139 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
140 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
141 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
142 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
143 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
144 const int8x16_t vbGHx01234567 = vshlq_n_s8(vbGHx0123456789ABCDEF, 4);
145 const int8x16_t vbIJx01234567 = vshlq_n_s8(vbIJx0123456789ABCDEF, 4);
146 const int8x16_t vbKLx01234567 = vshlq_n_s8(vbKLx0123456789ABCDEF, 4);
147 const int8x16_t vbMNx01234567 = vshlq_n_s8(vbMNx0123456789ABCDEF, 4);
148 const int8x16_t vbOPx01234567 = vshlq_n_s8(vbOPx0123456789ABCDEF, 4);
149 const int8x16_t vbQRx01234567 = vshlq_n_s8(vbQRx0123456789ABCDEF, 4);
150 const int8x16_t vbSTx01234567 = vshlq_n_s8(vbSTx0123456789ABCDEF, 4);
151 const int8x16_t vbUVx01234567 = vshlq_n_s8(vbUVx0123456789ABCDEF, 4);
152 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
153 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
154 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
155 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
156 const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask);
157 const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask);
158 const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask);
159 const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask);
160 const int8x16_t vbGHx89ABCDEF = vandq_s8(vbGHx0123456789ABCDEF, vmask);
161 const int8x16_t vbIJx89ABCDEF = vandq_s8(vbIJx0123456789ABCDEF, vmask);
162 const int8x16_t vbKLx89ABCDEF = vandq_s8(vbKLx0123456789ABCDEF, vmask);
163 const int8x16_t vbMNx89ABCDEF = vandq_s8(vbMNx0123456789ABCDEF, vmask);
164 const int8x16_t vbOPx89ABCDEF = vandq_s8(vbOPx0123456789ABCDEF, vmask);
165 const int8x16_t vbQRx89ABCDEF = vandq_s8(vbQRx0123456789ABCDEF, vmask);
166 const int8x16_t vbSTx89ABCDEF = vandq_s8(vbSTx0123456789ABCDEF, vmask);
167 const int8x16_t vbUVx89ABCDEF = vandq_s8(vbUVx0123456789ABCDEF, vmask);
168
169 // Multiply-accumulate: 2x8 * 8x32 --> 2x32.
170 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
171 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
172 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
173 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
174 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
175 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
176 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
177 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
178 vacc01xGH = vmmlaq_s32(vacc01xGH, va01x01234567, vbGHx01234567);
179 vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x01234567, vbIJx01234567);
180 vacc01xKL = vmmlaq_s32(vacc01xKL, va01x01234567, vbKLx01234567);
181 vacc01xMN = vmmlaq_s32(vacc01xMN, va01x01234567, vbMNx01234567);
182 vacc01xOP = vmmlaq_s32(vacc01xOP, va01x01234567, vbOPx01234567);
183 vacc01xQR = vmmlaq_s32(vacc01xQR, va01x01234567, vbQRx01234567);
184 vacc01xST = vmmlaq_s32(vacc01xST, va01x01234567, vbSTx01234567);
185 vacc01xUV = vmmlaq_s32(vacc01xUV, va01x01234567, vbUVx01234567);
186 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
187 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
188 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
189 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
190 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
191 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
192 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
193 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
194 vacc01xGH = vmmlaq_s32(vacc01xGH, va01x89ABCDEF, vbGHx89ABCDEF);
195 vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x89ABCDEF, vbIJx89ABCDEF);
196 vacc01xKL = vmmlaq_s32(vacc01xKL, va01x89ABCDEF, vbKLx89ABCDEF);
197 vacc01xMN = vmmlaq_s32(vacc01xMN, va01x89ABCDEF, vbMNx89ABCDEF);
198 vacc01xOP = vmmlaq_s32(vacc01xOP, va01x89ABCDEF, vbOPx89ABCDEF);
199 vacc01xQR = vmmlaq_s32(vacc01xQR, va01x89ABCDEF, vbQRx89ABCDEF);
200 vacc01xST = vmmlaq_s32(vacc01xST, va01x89ABCDEF, vbSTx89ABCDEF);
201 vacc01xUV = vmmlaq_s32(vacc01xUV, va01x89ABCDEF, vbUVx89ABCDEF);
202
203 k -= 16 * sizeof(int8_t);
204 }
205 // Handle up to 8 final positions of `k`
206 if XNN_UNLIKELY(k != 0) {
207 // Load a 2x8 block of activations.
208 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
209 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
210
211 // Load a 16x32 block of weights.
212 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
213 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
214 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
215 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
216 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
217 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
218 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
219 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
220 const int8x16_t vbGHx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
221 const int8x16_t vbIJx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
222 const int8x16_t vbKLx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
223 const int8x16_t vbMNx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
224 const int8x16_t vbOPx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
225 const int8x16_t vbQRx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
226 const int8x16_t vbSTx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
227 const int8x16_t vbUVx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
228 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
229 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
230 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
231 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
232 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
233 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
234 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
235 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
236 const int8x16_t vbGHx01234567 = vshlq_n_s8(vbGHx0123456789ABCDEF, 4);
237 const int8x16_t vbIJx01234567 = vshlq_n_s8(vbIJx0123456789ABCDEF, 4);
238 const int8x16_t vbKLx01234567 = vshlq_n_s8(vbKLx0123456789ABCDEF, 4);
239 const int8x16_t vbMNx01234567 = vshlq_n_s8(vbMNx0123456789ABCDEF, 4);
240 const int8x16_t vbOPx01234567 = vshlq_n_s8(vbOPx0123456789ABCDEF, 4);
241 const int8x16_t vbQRx01234567 = vshlq_n_s8(vbQRx0123456789ABCDEF, 4);
242 const int8x16_t vbSTx01234567 = vshlq_n_s8(vbSTx0123456789ABCDEF, 4);
243 const int8x16_t vbUVx01234567 = vshlq_n_s8(vbUVx0123456789ABCDEF, 4);
244
245 // Multiply-accumulate: 2x4 * 4x32 --> 2x32.
246 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
247 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
248 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
249 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
250 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
251 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
252 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
253 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
254 vacc01xGH = vmmlaq_s32(vacc01xGH, vreinterpretq_s8_u64(va01x01234567), vbGHx01234567);
255 vacc01xIJ = vmmlaq_s32(vacc01xIJ, vreinterpretq_s8_u64(va01x01234567), vbIJx01234567);
256 vacc01xKL = vmmlaq_s32(vacc01xKL, vreinterpretq_s8_u64(va01x01234567), vbKLx01234567);
257 vacc01xMN = vmmlaq_s32(vacc01xMN, vreinterpretq_s8_u64(va01x01234567), vbMNx01234567);
258 vacc01xOP = vmmlaq_s32(vacc01xOP, vreinterpretq_s8_u64(va01x01234567), vbOPx01234567);
259 vacc01xQR = vmmlaq_s32(vacc01xQR, vreinterpretq_s8_u64(va01x01234567), vbQRx01234567);
260 vacc01xST = vmmlaq_s32(vacc01xST, vreinterpretq_s8_u64(va01x01234567), vbSTx01234567);
261 vacc01xUV = vmmlaq_s32(vacc01xUV, vreinterpretq_s8_u64(va01x01234567), vbUVx01234567);
262 }
263
264 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
265 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
266 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
267 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
268 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
269 int32x4_t vacc1x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
270 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
271 int32x4_t vacc1xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
272 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xGH), vreinterpretq_u64_s32(vacc01xIJ)));
273 int32x4_t vacc1xGHIJ = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xGH), vreinterpretq_u64_s32(vacc01xIJ)));
274 int32x4_t vacc0xKLMN = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xKL), vreinterpretq_u64_s32(vacc01xMN)));
275 int32x4_t vacc1xKLMN = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xKL), vreinterpretq_u64_s32(vacc01xMN)));
276 int32x4_t vacc0xOPQR = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xOP), vreinterpretq_u64_s32(vacc01xQR)));
277 int32x4_t vacc1xOPQR = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xOP), vreinterpretq_u64_s32(vacc01xQR)));
278 int32x4_t vacc0xSTUV = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xST), vreinterpretq_u64_s32(vacc01xUV)));
279 int32x4_t vacc1xSTUV = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xST), vreinterpretq_u64_s32(vacc01xUV)));
280 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
281 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
282 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
283 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
284 const float32x4_t vfilter_output_scaleGHIJ = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
285 const float32x4_t vfilter_output_scaleKLMN = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
286 const float32x4_t vfilter_output_scaleOPQR = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
287 const float32x4_t vfilter_output_scaleSTUV = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
288 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
289 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
290 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
291 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
292 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
293 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
294 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
295 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
296 float32x4_t vf0xGHIJ = vcvtq_f32_s32(vacc0xGHIJ);
297 vout0xGHIJ = vfmaq_f32(vout0xGHIJ, vf0xGHIJ, vfilter_output_scaleGHIJ);
298 float32x4_t vf0xKLMN = vcvtq_f32_s32(vacc0xKLMN);
299 vout0xKLMN = vfmaq_f32(vout0xKLMN, vf0xKLMN, vfilter_output_scaleKLMN);
300 float32x4_t vf0xOPQR = vcvtq_f32_s32(vacc0xOPQR);
301 vout0xOPQR = vfmaq_f32(vout0xOPQR, vf0xOPQR, vfilter_output_scaleOPQR);
302 float32x4_t vf0xSTUV = vcvtq_f32_s32(vacc0xSTUV);
303 vout0xSTUV = vfmaq_f32(vout0xSTUV, vf0xSTUV, vfilter_output_scaleSTUV);
304 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
305 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
306 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
307 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
308 float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
309 vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
310 float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
311 vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
312 float32x4_t vf1xGHIJ = vcvtq_f32_s32(vacc1xGHIJ);
313 vout1xGHIJ = vfmaq_f32(vout1xGHIJ, vf1xGHIJ, vfilter_output_scaleGHIJ);
314 float32x4_t vf1xKLMN = vcvtq_f32_s32(vacc1xKLMN);
315 vout1xKLMN = vfmaq_f32(vout1xKLMN, vf1xKLMN, vfilter_output_scaleKLMN);
316 float32x4_t vf1xOPQR = vcvtq_f32_s32(vacc1xOPQR);
317 vout1xOPQR = vfmaq_f32(vout1xOPQR, vf1xOPQR, vfilter_output_scaleOPQR);
318 float32x4_t vf1xSTUV = vcvtq_f32_s32(vacc1xSTUV);
319 vout1xSTUV = vfmaq_f32(vout1xSTUV, vf1xSTUV, vfilter_output_scaleSTUV);
320 }
321
322 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
323 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
324 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
325 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
326 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
327 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
328 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
329 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
330 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
331 vout0xGHIJ = vmulq_lane_f32(vout0xGHIJ, vget_low_f32(vinput_scale01), 1);
332 vout1xGHIJ = vmulq_lane_f32(vout1xGHIJ, vget_high_f32(vinput_scale01), 1);
333 vout0xKLMN = vmulq_lane_f32(vout0xKLMN, vget_low_f32(vinput_scale01), 1);
334 vout1xKLMN = vmulq_lane_f32(vout1xKLMN, vget_high_f32(vinput_scale01), 1);
335 vout0xOPQR = vmulq_lane_f32(vout0xOPQR, vget_low_f32(vinput_scale01), 1);
336 vout1xOPQR = vmulq_lane_f32(vout1xOPQR, vget_high_f32(vinput_scale01), 1);
337 vout0xSTUV = vmulq_lane_f32(vout0xSTUV, vget_low_f32(vinput_scale01), 1);
338 vout1xSTUV = vmulq_lane_f32(vout1xSTUV, vget_high_f32(vinput_scale01), 1);
339
340
341 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
342 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
343 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
344 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
345 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
346 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
347 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
348 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
349 vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
350 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
351 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
352 vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);
353 const float32x4_t vbiasGHIJ = vld1q_f32(w); w = (const float*) w + 4;
354 vout0xGHIJ = vaddq_f32(vbiasGHIJ, vout0xGHIJ);
355 vout1xGHIJ = vaddq_f32(vbiasGHIJ, vout1xGHIJ);
356 const float32x4_t vbiasKLMN = vld1q_f32(w); w = (const float*) w + 4;
357 vout0xKLMN = vaddq_f32(vbiasKLMN, vout0xKLMN);
358 vout1xKLMN = vaddq_f32(vbiasKLMN, vout1xKLMN);
359 const float32x4_t vbiasOPQR = vld1q_f32(w); w = (const float*) w + 4;
360 vout0xOPQR = vaddq_f32(vbiasOPQR, vout0xOPQR);
361 vout1xOPQR = vaddq_f32(vbiasOPQR, vout1xOPQR);
362 const float32x4_t vbiasSTUV = vld1q_f32(w); w = (const float*) w + 4;
363 vout0xSTUV = vaddq_f32(vbiasSTUV, vout0xSTUV);
364 vout1xSTUV = vaddq_f32(vbiasSTUV, vout1xSTUV);
365
366 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
367 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
368 float16x8_t vfp16out0xGHIJKLMN = vcombine_f16(vcvt_f16_f32(vout0xGHIJ), vcvt_f16_f32(vout0xKLMN));
369 float16x8_t vfp16out0xOPQRSTUV = vcombine_f16(vcvt_f16_f32(vout0xOPQR), vcvt_f16_f32(vout0xSTUV));
370 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
371 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
372 float16x8_t vfp16out1xGHIJKLMN = vcombine_f16(vcvt_f16_f32(vout1xGHIJ), vcvt_f16_f32(vout1xKLMN));
373 float16x8_t vfp16out1xOPQRSTUV = vcombine_f16(vcvt_f16_f32(vout1xOPQR), vcvt_f16_f32(vout1xSTUV));
374
375 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
376 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
377 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
378 vfp16out0xGHIJKLMN = vmaxq_f16(vfp16out0xGHIJKLMN, voutput_min);
379 vfp16out0xOPQRSTUV = vmaxq_f16(vfp16out0xOPQRSTUV, voutput_min);
380 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
381 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
382 vfp16out1xGHIJKLMN = vmaxq_f16(vfp16out1xGHIJKLMN, voutput_min);
383 vfp16out1xOPQRSTUV = vmaxq_f16(vfp16out1xOPQRSTUV, voutput_min);
384 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
385 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
386 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
387 vfp16out0xGHIJKLMN = vminq_f16(vfp16out0xGHIJKLMN, voutput_max);
388 vfp16out0xOPQRSTUV = vminq_f16(vfp16out0xOPQRSTUV, voutput_max);
389 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
390 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
391 vfp16out1xGHIJKLMN = vminq_f16(vfp16out1xGHIJKLMN, voutput_max);
392 vfp16out1xOPQRSTUV = vminq_f16(vfp16out1xOPQRSTUV, voutput_max);
393 if XNN_LIKELY(nc >= 32) {
394 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
395 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
396 vst1q_u16(c0 + 16, vreinterpretq_u16_f16(vfp16out0xGHIJKLMN));
397 vst1q_u16(c0 + 24, vreinterpretq_u16_f16(vfp16out0xOPQRSTUV));
398 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
399 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
400 vst1q_u16(c1 + 16, vreinterpretq_u16_f16(vfp16out1xGHIJKLMN));
401 vst1q_u16(c1 + 24, vreinterpretq_u16_f16(vfp16out1xOPQRSTUV));
402
403 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
404 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
405
406 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
407 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
408
409 nc -= 32;
410 } else {
411 if (nc & 16) {
412 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
413 vfp16out0x01234567 = vfp16out0xGHIJKLMN;
414 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
415 vfp16out1x01234567 = vfp16out1xGHIJKLMN;
416 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x89ABCDEF)); c0 += 8;
417 vfp16out0x89ABCDEF = vfp16out0xOPQRSTUV;
418 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x89ABCDEF)); c1 += 8;
419 vfp16out1x89ABCDEF = vfp16out1xOPQRSTUV;
420 }
421 if (nc & 8) {
422 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
423 vfp16out0x01234567 = vfp16out0x89ABCDEF;
424 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
425 vfp16out1x01234567 = vfp16out1x89ABCDEF;
426 }
427 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
428 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
429 if (nc & 4) {
430 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
431 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
432 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
433 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
434 }
435 if (nc & 2) {
436 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
437 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
438 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
439 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
440 }
441 if (nc & 1) {
442 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
443 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
444 }
445 nc = 0;
446 }
447 } while (nc != 0);
448 }