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_gemminc_minmax_ukernel_4x8__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 void * restrict acc,const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_gemminc_minmax_ukernel_4x8__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 void*restrict acc,
31 const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
32 {
33 assert(mr != 0);
34 assert(mr <= 4);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(__fp16) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41 assert(acc != NULL);
42
43 const __fp16* a0 = (const __fp16*) a;
44 __fp16* c0 = (__fp16*) c;
45 const __fp16* a1 = (const __fp16*) ((uintptr_t) a0 + a_stride);
46 __fp16* c1 = (__fp16*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const __fp16* a2 = (const __fp16*) ((uintptr_t) a1 + a_stride);
52 __fp16* c2 = (__fp16*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57 const __fp16* a3 = (const __fp16*) ((uintptr_t) a2 + a_stride);
58 __fp16* c3 = (__fp16*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr != 4) {
60 a3 = a2;
61 c3 = c2;
62 }
63
64 do {
65 float16x8_t vacc0x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
66 float16x8_t vacc1x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
67 float16x8_t vacc2x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
68 float16x8_t vacc3x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
69
70 size_t k = kc;
71 while (k >= 4 * sizeof(__fp16)) {
72 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
73 const float16x4_t va1 = vld1_f16(a1); a1 += 4;
74 const float16x4_t va2 = vld1_f16(a2); a2 += 4;
75 const float16x4_t va3 = vld1_f16(a3); a3 += 4;
76
77 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
78
79 #if XNN_ARCH_ARM64
80 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
81 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
82 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
83 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
84 #else
85 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
86 const float16x8_t va1c0 = vdupq_lane_f16(va1, 0);
87 const float16x8_t va2c0 = vdupq_lane_f16(va2, 0);
88 const float16x8_t va3c0 = vdupq_lane_f16(va3, 0);
89
90 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
91 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c0, vb01234567c0);
92 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);
93 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c0, vb01234567c0);
94 #endif
95 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
96
97 #if XNN_ARCH_ARM64
98 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
99 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
100 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
101 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
102 #else
103 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
104 const float16x8_t va1c1 = vdupq_lane_f16(va1, 1);
105 const float16x8_t va2c1 = vdupq_lane_f16(va2, 1);
106 const float16x8_t va3c1 = vdupq_lane_f16(va3, 1);
107
108 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
109 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c1, vb01234567c1);
110 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);
111 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c1, vb01234567c1);
112 #endif
113 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
114
115 #if XNN_ARCH_ARM64
116 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
117 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
118 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
119 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
120 #else
121 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
122 const float16x8_t va1c2 = vdupq_lane_f16(va1, 2);
123 const float16x8_t va2c2 = vdupq_lane_f16(va2, 2);
124 const float16x8_t va3c2 = vdupq_lane_f16(va3, 2);
125
126 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
127 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c2, vb01234567c2);
128 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);
129 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c2, vb01234567c2);
130 #endif
131 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
132
133 #if XNN_ARCH_ARM64
134 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
135 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
136 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
137 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
138 #else
139 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
140 const float16x8_t va1c3 = vdupq_lane_f16(va1, 3);
141 const float16x8_t va2c3 = vdupq_lane_f16(va2, 3);
142 const float16x8_t va3c3 = vdupq_lane_f16(va3, 3);
143
144 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
145 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c3, vb01234567c3);
146 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);
147 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c3, vb01234567c3);
148 #endif
149
150 k -= 4 * sizeof(__fp16);
151 }
152 if XNN_UNLIKELY(k != 0) {
153 do {
154 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
155 const float16x8_t va1 = vld1q_dup_f16(a1); a1 += 1;
156 const float16x8_t va2 = vld1q_dup_f16(a2); a2 += 1;
157 const float16x8_t va3 = vld1q_dup_f16(a3); a3 += 1;
158
159 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
160
161 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
162 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
163 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
164 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
165
166 k -= sizeof(__fp16);
167 } while (k != 0);
168 }
169
170 const float16x8_t vscale = vld1q_dup_f16((const __fp16*) ¶ms->scale);
171 vacc0x01234567 = vmulq_f16(vacc0x01234567, vscale);
172 vacc1x01234567 = vmulq_f16(vacc1x01234567, vscale);
173 vacc2x01234567 = vmulq_f16(vacc2x01234567, vscale);
174 vacc3x01234567 = vmulq_f16(vacc3x01234567, vscale);
175
176 const float16x8_t vmax = vld1q_dup_f16((const __fp16*) ¶ms->max);
177 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
178 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
179 vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
180 vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
181
182 const float16x8_t vmin = vld1q_dup_f16((const __fp16*) ¶ms->min);
183 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
184 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
185 vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
186 vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
187
188 if XNN_LIKELY(nc >= 8) {
189 vst1q_f16(c0, vacc0x01234567);
190 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
191 vst1q_f16(c1, vacc1x01234567);
192 c1 = (__fp16*) ((uintptr_t) c1 + cn_stride);
193 vst1q_f16(c2, vacc2x01234567);
194 c2 = (__fp16*) ((uintptr_t) c2 + cn_stride);
195 vst1q_f16(c3, vacc3x01234567);
196 c3 = (__fp16*) ((uintptr_t) c3 + cn_stride);
197
198 a0 = (const __fp16*) ((uintptr_t) a0 - kc);
199 a1 = (const __fp16*) ((uintptr_t) a1 - kc);
200 a2 = (const __fp16*) ((uintptr_t) a2 - kc);
201 a3 = (const __fp16*) ((uintptr_t) a3 - kc);
202
203 nc -= 8;
204 } else {
205 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
206 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
207 float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
208 float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
209 if (nc & 4) {
210 vst1_f16(c0, vacc0x0123); c0 += 4;
211 vst1_f16(c1, vacc1x0123); c1 += 4;
212 vst1_f16(c2, vacc2x0123); c2 += 4;
213 vst1_f16(c3, vacc3x0123); c3 += 4;
214
215 vacc0x0123 = vget_high_f16(vacc0x01234567);
216 vacc1x0123 = vget_high_f16(vacc1x01234567);
217 vacc2x0123 = vget_high_f16(vacc2x01234567);
218 vacc3x0123 = vget_high_f16(vacc3x01234567);
219 }
220 if (nc & 2) {
221 vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
222 vst1_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
223 vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
224 vst1_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
225
226 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
227 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
228 vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
229 vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
230 }
231 if (nc & 1) {
232 vst1_lane_f16(c0, vacc0x0123, 0);
233 vst1_lane_f16(c1, vacc1x0123, 0);
234 vst1_lane_f16(c2, vacc2x0123, 0);
235 vst1_lane_f16(c3, vacc3x0123, 0);
236 }
237
238 nc = 0;
239 }
240 } while (nc != 0);
241 }
242