1 // Auto-generated file. Do not edit!
2 // Template: src/f16-gemm/neonfp16arith-ld64.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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include <xnnpack/common.h>
16
17 #include <xnnpack/gemm.h>
18
19
xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const void* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 void* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(__fp16) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 const __fp16* a0 = (const __fp16*) a;
42 __fp16* c0 = (__fp16*) c;
43 const __fp16* a1 = (const __fp16*) ((uintptr_t) a0 + a_stride);
44 __fp16* c1 = (__fp16*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const __fp16* a2 = (const __fp16*) ((uintptr_t) a1 + a_stride);
50 __fp16* c2 = (__fp16*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const __fp16* a3 = (const __fp16*) ((uintptr_t) a2 + a_stride);
56 __fp16* c3 = (__fp16*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr != 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61
62 do {
63 float16x8_t vacc0x01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
64 float16x8_t vacc0x89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
65 float16x8_t vacc1x01234567 = vacc0x01234567;
66 float16x8_t vacc1x89ABCDEF = vacc0x89ABCDEF;
67 float16x8_t vacc2x01234567 = vacc0x01234567;
68 float16x8_t vacc2x89ABCDEF = vacc0x89ABCDEF;
69 float16x8_t vacc3x01234567 = vacc0x01234567;
70 float16x8_t vacc3x89ABCDEF = vacc0x89ABCDEF;
71
72 size_t k = kc;
73 while (k >= 4 * sizeof(__fp16)) {
74 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
75 const float16x4_t va1 = vld1_f16(a1); a1 += 4;
76 const float16x4_t va2 = vld1_f16(a2); a2 += 4;
77 const float16x4_t va3 = vld1_f16(a3); a3 += 4;
78
79 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
80 const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
81
82 #if XNN_ARCH_ARM64
83 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
84 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
85 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
86 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
87 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
88 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc0, va1, 0);
89 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc0, va2, 0);
90 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc0, va3, 0);
91 #else
92 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
93 const float16x8_t va1c0 = vdupq_lane_f16(va1, 0);
94 const float16x8_t va2c0 = vdupq_lane_f16(va2, 0);
95 const float16x8_t va3c0 = vdupq_lane_f16(va3, 0);
96
97 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
98 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c0, vb01234567c0);
99 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);
100 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c0, vb01234567c0);
101 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c0, vb89ABCDEFc0);
102 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c0, vb89ABCDEFc0);
103 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c0, vb89ABCDEFc0);
104 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c0, vb89ABCDEFc0);
105 #endif
106 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
107 const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
108
109 #if XNN_ARCH_ARM64
110 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
111 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
112 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
113 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
114 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
115 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc1, va1, 1);
116 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc1, va2, 1);
117 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc1, va3, 1);
118 #else
119 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
120 const float16x8_t va1c1 = vdupq_lane_f16(va1, 1);
121 const float16x8_t va2c1 = vdupq_lane_f16(va2, 1);
122 const float16x8_t va3c1 = vdupq_lane_f16(va3, 1);
123
124 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
125 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c1, vb01234567c1);
126 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);
127 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c1, vb01234567c1);
128 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c1, vb89ABCDEFc1);
129 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c1, vb89ABCDEFc1);
130 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c1, vb89ABCDEFc1);
131 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c1, vb89ABCDEFc1);
132 #endif
133 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
134 const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
135
136 #if XNN_ARCH_ARM64
137 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
138 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
139 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
140 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
141 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
142 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc2, va1, 2);
143 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc2, va2, 2);
144 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc2, va3, 2);
145 #else
146 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
147 const float16x8_t va1c2 = vdupq_lane_f16(va1, 2);
148 const float16x8_t va2c2 = vdupq_lane_f16(va2, 2);
149 const float16x8_t va3c2 = vdupq_lane_f16(va3, 2);
150
151 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
152 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c2, vb01234567c2);
153 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);
154 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c2, vb01234567c2);
155 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c2, vb89ABCDEFc2);
156 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c2, vb89ABCDEFc2);
157 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c2, vb89ABCDEFc2);
158 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c2, vb89ABCDEFc2);
159 #endif
160 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
161 const float16x8_t vb89ABCDEFc3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
162
163 #if XNN_ARCH_ARM64
164 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
165 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
166 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
167 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
168 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
169 vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc3, va1, 3);
170 vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc3, va2, 3);
171 vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc3, va3, 3);
172 #else
173 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
174 const float16x8_t va1c3 = vdupq_lane_f16(va1, 3);
175 const float16x8_t va2c3 = vdupq_lane_f16(va2, 3);
176 const float16x8_t va3c3 = vdupq_lane_f16(va3, 3);
177
178 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
179 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c3, vb01234567c3);
180 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);
181 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c3, vb01234567c3);
182 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c3, vb89ABCDEFc3);
183 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c3, vb89ABCDEFc3);
184 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c3, vb89ABCDEFc3);
185 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c3, vb89ABCDEFc3);
186 #endif
187
188 k -= 4 * sizeof(__fp16);
189 }
190 if XNN_UNLIKELY(k != 0) {
191 do {
192 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
193 const float16x8_t va1 = vld1q_dup_f16(a1); a1 += 1;
194 const float16x8_t va2 = vld1q_dup_f16(a2); a2 += 1;
195 const float16x8_t va3 = vld1q_dup_f16(a3); a3 += 1;
196
197 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
198 const float16x8_t vb89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
199
200 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
201 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
202 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
203 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
204 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);
205 vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1, vb89ABCDEF);
206 vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2, vb89ABCDEF);
207 vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3, vb89ABCDEF);
208
209 k -= sizeof(__fp16);
210 } while (k != 0);
211 }
212
213
214 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
215 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
216 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
217 vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
218 vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
219 vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
220 vacc1x89ABCDEF = vminq_f16(vacc1x89ABCDEF, vmax);
221 vacc2x89ABCDEF = vminq_f16(vacc2x89ABCDEF, vmax);
222 vacc3x89ABCDEF = vminq_f16(vacc3x89ABCDEF, vmax);
223
224 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
225 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
226 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
227 vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
228 vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
229 vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
230 vacc1x89ABCDEF = vmaxq_f16(vacc1x89ABCDEF, vmin);
231 vacc2x89ABCDEF = vmaxq_f16(vacc2x89ABCDEF, vmin);
232 vacc3x89ABCDEF = vmaxq_f16(vacc3x89ABCDEF, vmin);
233
234 if XNN_LIKELY(nc >= 16) {
235 vst1q_f16(c0, vacc0x01234567);
236 vst1q_f16(c0 + 8, vacc0x89ABCDEF);
237 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
238 vst1q_f16(c1, vacc1x01234567);
239 vst1q_f16(c1 + 8, vacc1x89ABCDEF);
240 c1 = (__fp16*) ((uintptr_t) c1 + cn_stride);
241 vst1q_f16(c2, vacc2x01234567);
242 vst1q_f16(c2 + 8, vacc2x89ABCDEF);
243 c2 = (__fp16*) ((uintptr_t) c2 + cn_stride);
244 vst1q_f16(c3, vacc3x01234567);
245 vst1q_f16(c3 + 8, vacc3x89ABCDEF);
246 c3 = (__fp16*) ((uintptr_t) c3 + cn_stride);
247
248 a0 = (const __fp16*) ((uintptr_t) a0 - kc);
249 a1 = (const __fp16*) ((uintptr_t) a1 - kc);
250 a2 = (const __fp16*) ((uintptr_t) a2 - kc);
251 a3 = (const __fp16*) ((uintptr_t) a3 - kc);
252
253 nc -= 16;
254 } else {
255 if (nc & 8) {
256 vst1q_f16(c0, vacc0x01234567); c0 += 8;
257 vst1q_f16(c1, vacc1x01234567); c1 += 8;
258 vst1q_f16(c2, vacc2x01234567); c2 += 8;
259 vst1q_f16(c3, vacc3x01234567); c3 += 8;
260
261 vacc0x01234567 = vacc0x89ABCDEF;
262 vacc1x01234567 = vacc1x89ABCDEF;
263 vacc2x01234567 = vacc2x89ABCDEF;
264 vacc3x01234567 = vacc3x89ABCDEF;
265 }
266 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
267 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
268 float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
269 float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
270 if (nc & 4) {
271 vst1_f16(c0, vacc0x0123); c0 += 4;
272 vst1_f16(c1, vacc1x0123); c1 += 4;
273 vst1_f16(c2, vacc2x0123); c2 += 4;
274 vst1_f16(c3, vacc3x0123); c3 += 4;
275
276 vacc0x0123 = vget_high_f16(vacc0x01234567);
277 vacc1x0123 = vget_high_f16(vacc1x01234567);
278 vacc2x0123 = vget_high_f16(vacc2x01234567);
279 vacc3x0123 = vget_high_f16(vacc3x01234567);
280 }
281 if (nc & 2) {
282 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
283 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
284 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
285 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
286
287 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
288 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
289 vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
290 vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
291 }
292 if (nc & 1) {
293 vst1_lane_f16(c0, vacc0x0123, 0);
294 vst1_lane_f16(c1, vacc1x0123, 0);
295 vst1_lane_f16(c2, vacc2x0123, 0);
296 vst1_lane_f16(c3, vacc3x0123, 0);
297 }
298
299 nc = 0;
300 }
301 } while (nc != 0);
302 }
303