1 // Auto-generated file. Do not edit!
2 // Template: src/f16-igemm/neonfp16arith-ld64.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/igemm.h>
16
17
xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const void ** restrict a,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const void * zero,const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const void** restrict a,
24 const void* restrict w,
25 void* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const void* zero,
30 const union xnn_f16_scaleminmax_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(ks != 0);
38 assert(ks % (4 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(__fp16) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 __fp16* c0 = (__fp16*) c;
45 __fp16* c1 = (__fp16*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 __fp16* c2 = (__fp16*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 __fp16* c3 = (__fp16*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 do {
59 float16x8_t vacc0x01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
60 float16x8_t vacc1x01234567 = vacc0x01234567;
61 float16x8_t vacc2x01234567 = vacc0x01234567;
62 float16x8_t vacc3x01234567 = vacc0x01234567;
63
64 size_t p = ks;
65 do {
66 const __fp16* restrict a0 = (const __fp16*) a[0];
67 assert(a0 != NULL);
68 if XNN_UNPREDICTABLE(a0 != zero) {
69 a0 = (const __fp16*) ((uintptr_t) a0 + a_offset);
70 }
71 const __fp16* restrict a1 = (const __fp16*) a[1];
72 assert(a1 != NULL);
73 if XNN_UNPREDICTABLE(a1 != zero) {
74 a1 = (const __fp16*) ((uintptr_t) a1 + a_offset);
75 }
76 const __fp16* restrict a2 = (const __fp16*) a[2];
77 assert(a2 != NULL);
78 if XNN_UNPREDICTABLE(a2 != zero) {
79 a2 = (const __fp16*) ((uintptr_t) a2 + a_offset);
80 }
81 const __fp16* restrict a3 = (const __fp16*) a[3];
82 assert(a3 != NULL);
83 if XNN_UNPREDICTABLE(a3 != zero) {
84 a3 = (const __fp16*) ((uintptr_t) a3 + a_offset);
85 }
86 a += 4;
87
88 size_t k = kc;
89 for (; k >= 4 * sizeof(__fp16); k -= 4 * sizeof(__fp16)) {
90 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
91 const float16x4_t va1 = vld1_f16(a1); a1 += 4;
92 const float16x4_t va2 = vld1_f16(a2); a2 += 4;
93 const float16x4_t va3 = vld1_f16(a3); a3 += 4;
94
95 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
96
97 #if XNN_ARCH_ARM64
98 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
99 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
100 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
101 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
102 #else
103 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
104 const float16x8_t va1c0 = vdupq_lane_f16(va1, 0);
105 const float16x8_t va2c0 = vdupq_lane_f16(va2, 0);
106 const float16x8_t va3c0 = vdupq_lane_f16(va3, 0);
107
108 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
109 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c0, vb01234567c0);
110 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);
111 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c0, vb01234567c0);
112 #endif
113 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
114
115 #if XNN_ARCH_ARM64
116 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
117 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
118 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
119 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
120 #else
121 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
122 const float16x8_t va1c1 = vdupq_lane_f16(va1, 1);
123 const float16x8_t va2c1 = vdupq_lane_f16(va2, 1);
124 const float16x8_t va3c1 = vdupq_lane_f16(va3, 1);
125
126 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
127 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c1, vb01234567c1);
128 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);
129 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c1, vb01234567c1);
130 #endif
131 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
132
133 #if XNN_ARCH_ARM64
134 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
135 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
136 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
137 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
138 #else
139 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
140 const float16x8_t va1c2 = vdupq_lane_f16(va1, 2);
141 const float16x8_t va2c2 = vdupq_lane_f16(va2, 2);
142 const float16x8_t va3c2 = vdupq_lane_f16(va3, 2);
143
144 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
145 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c2, vb01234567c2);
146 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);
147 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c2, vb01234567c2);
148 #endif
149 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
150
151 #if XNN_ARCH_ARM64
152 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
153 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
154 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
155 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
156 #else
157 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
158 const float16x8_t va1c3 = vdupq_lane_f16(va1, 3);
159 const float16x8_t va2c3 = vdupq_lane_f16(va2, 3);
160 const float16x8_t va3c3 = vdupq_lane_f16(va3, 3);
161
162 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
163 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c3, vb01234567c3);
164 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);
165 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c3, vb01234567c3);
166 #endif
167 }
168 if XNN_UNLIKELY(k != 0) {
169 do {
170 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
171 const float16x8_t va1 = vld1q_dup_f16(a1); a1 += 1;
172 const float16x8_t va2 = vld1q_dup_f16(a2); a2 += 1;
173 const float16x8_t va3 = vld1q_dup_f16(a3); a3 += 1;
174
175 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
176
177 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
178 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
179 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
180 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
181
182 k -= sizeof(__fp16);
183 } while (k != 0);
184 }
185 p -= 4 * sizeof(void*);
186 } while (p != 0);
187
188 const float16x8_t vscale = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.scale));
189 vacc0x01234567 = vmulq_f16(vacc0x01234567, vscale);
190 vacc1x01234567 = vmulq_f16(vacc1x01234567, vscale);
191 vacc2x01234567 = vmulq_f16(vacc2x01234567, vscale);
192 vacc3x01234567 = vmulq_f16(vacc3x01234567, vscale);
193
194 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
195 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
196 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
197 vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
198 vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
199
200 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
201 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
202 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
203 vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
204 vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
205
206 if XNN_LIKELY(nc >= 8) {
207 vst1q_f16(c3, vacc3x01234567);
208 c3 = (__fp16*) ((uintptr_t) c3 + cn_stride);
209 vst1q_f16(c2, vacc2x01234567);
210 c2 = (__fp16*) ((uintptr_t) c2 + cn_stride);
211 vst1q_f16(c1, vacc1x01234567);
212 c1 = (__fp16*) ((uintptr_t) c1 + cn_stride);
213 vst1q_f16(c0, vacc0x01234567);
214 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
215
216 a = (const void**restrict) ((uintptr_t) a - ks);
217 nc -= 8;
218 } else {
219 float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
220 float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
221 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
222 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
223 if (nc & 4) {
224 vst1_f16(c3, vacc3x0123); c3 += 4;
225 vst1_f16(c2, vacc2x0123); c2 += 4;
226 vst1_f16(c1, vacc1x0123); c1 += 4;
227 vst1_f16(c0, vacc0x0123); c0 += 4;
228
229 vacc3x0123 = vget_high_f16(vacc3x01234567);
230 vacc2x0123 = vget_high_f16(vacc2x01234567);
231 vacc1x0123 = vget_high_f16(vacc1x01234567);
232 vacc0x0123 = vget_high_f16(vacc0x01234567);
233 }
234 if (nc & 2) {
235 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
236 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
237 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
238 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
239
240 vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
241 vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
242 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
243 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
244 }
245 if (nc & 1) {
246 vst1_lane_f16(c3, vacc3x0123, 0);
247 vst1_lane_f16(c2, vacc2x0123, 0);
248 vst1_lane_f16(c1, vacc1x0123, 0);
249 vst1_lane_f16(c0, vacc0x0123, 0);
250 }
251
252 nc = 0;
253 }
254 } while (nc != 0);
255 }
256