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_1x16__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 struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_igemm_minmax_ukernel_1x16__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 struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(__fp16) == 0);
37 assert(ks != 0);
38 assert(ks % (1 * 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
46 do {
47 float16x8_t vacc0x01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
48 float16x8_t vacc0x89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
49
50 size_t p = ks;
51 do {
52 const __fp16* restrict a0 = (const __fp16*) a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const __fp16*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 for (; k >= 4 * sizeof(__fp16); k -= 4 * sizeof(__fp16)) {
61 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
62
63 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x8_t));
64 const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x8_t));
65
66 #if XNN_ARCH_ARM64
67 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
68 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
69 #else
70 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
71
72 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
73 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c0, vb89ABCDEFc0);
74 #endif
75 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x8_t));
76 const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x8_t));
77
78 #if XNN_ARCH_ARM64
79 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
80 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
81 #else
82 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
83
84 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
85 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c1, vb89ABCDEFc1);
86 #endif
87 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x8_t));
88 const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x8_t));
89
90 #if XNN_ARCH_ARM64
91 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
92 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
93 #else
94 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
95
96 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
97 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c2, vb89ABCDEFc2);
98 #endif
99 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x8_t));
100 const float16x8_t vb89ABCDEFc3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof( float16x8_t));
101
102 #if XNN_ARCH_ARM64
103 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
104 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
105 #else
106 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
107
108 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
109 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c3, vb89ABCDEFc3);
110 #endif
111 }
112 if XNN_UNLIKELY(k != 0) {
113 do {
114 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
115
116 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
117 const float16x8_t vb89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
118
119 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
120 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);
121
122 k -= sizeof(__fp16);
123 } while (k != 0);
124 }
125 p -= 1 * sizeof(void*);
126 } while (p != 0);
127
128 const float16x8_t vscale = vld1q_dup_f16((const __fp16*) ¶ms->scale);
129 vacc0x01234567 = vmulq_f16(vacc0x01234567, vscale);
130 vacc0x89ABCDEF = vmulq_f16(vacc0x89ABCDEF, vscale);
131
132 const float16x8_t vmax = vld1q_dup_f16((const __fp16*) ¶ms->max);
133 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
134 vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
135
136 const float16x8_t vmin = vld1q_dup_f16((const __fp16*) ¶ms->min);
137 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
138 vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
139
140 if XNN_LIKELY(nc >= 16) {
141 vst1q_f16(c0, vacc0x01234567);
142 vst1q_f16(c0 + 8, vacc0x89ABCDEF);
143 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
144
145 a = (const void**restrict) ((uintptr_t) a - ks);
146 nc -= 16;
147 } else {
148 if (nc & 8) {
149 vst1q_f16(c0, vacc0x01234567); c0 += 8;
150
151 vacc0x01234567 = vacc0x89ABCDEF;
152 }
153 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
154 if (nc & 4) {
155 vst1_f16(c0, vacc0x0123); c0 += 4;
156
157 vacc0x0123 = vget_high_f16(vacc0x01234567);
158 }
159 if (nc & 2) {
160 vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
161
162 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
163 }
164 if (nc & 1) {
165 vst1_lane_f16(c0, vacc0x0123, 0);
166 }
167
168 nc = 0;
169 }
170 } while (nc != 0);
171 }
172