1 // Auto-generated file. Do not edit!
2 // Template: src/f16-vmulcaddc/neonfp16arith.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 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/vmulcaddc.h>
16
17
xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x(size_t rows,size_t channels,const void * restrict input,size_t input_stride,const void * restrict weights,void * restrict output,size_t output_stride,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vmulcaddc_minmax_ukernel_c16__neonfp16arith_2x(
19 size_t rows,
20 size_t channels,
21 const void*restrict input,
22 size_t input_stride,
23 const void*restrict weights,
24 void*restrict output,
25 size_t output_stride,
26 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28 assert(rows != 0);
29 assert(channels != 0);
30 assert(channels % sizeof(__fp16) == 0);
31
32 const __fp16* i0 = (const __fp16*) input;
33 __fp16* o0 = (__fp16*) output;
34 const __fp16* i1 = (const __fp16*) ((uintptr_t) i0 + input_stride);
35 __fp16* o1 = (__fp16*) ((uintptr_t) o0 + output_stride);
36
37 const size_t input_increment = input_stride * 2 - channels;
38 const size_t output_increment = output_stride * 2 - channels;
39
40 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
41 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
42 do {
43 if XNN_UNPREDICTABLE(rows < 2) {
44 i1 = i0;
45 o1 = o0;
46 }
47
48 const __fp16* w = (const __fp16*) weights;
49 size_t c = channels;
50 for (; c >= 16 * sizeof(__fp16); c -= 16 * sizeof(__fp16)) {
51 const float16x8_t vscale01234567 = vld1q_f16(w); w += 8;
52 const float16x8_t vscale89ABCDEF = vld1q_f16(w); w += 8;
53
54 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 += 8;
55 float16x8_t vacc0x89ABCDEF = vld1q_f16(i0); i0 += 8;
56 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 += 8;
57 float16x8_t vacc1x89ABCDEF = vld1q_f16(i1); i1 += 8;
58
59 const float16x8_t vbias01234567 = vld1q_f16(w); w += 8;
60 const float16x8_t vbias89ABCDEF = vld1q_f16(w); w += 8;
61
62 vacc0x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc0x01234567);
63 vacc0x89ABCDEF = vfmaq_f16(vbias89ABCDEF, vscale89ABCDEF, vacc0x89ABCDEF);
64 vacc1x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc1x01234567);
65 vacc1x89ABCDEF = vfmaq_f16(vbias89ABCDEF, vscale89ABCDEF, vacc1x89ABCDEF);
66
67 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
68 vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
69 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
70 vacc1x89ABCDEF = vmaxq_f16(vacc1x89ABCDEF, vmin);
71
72 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
73 vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
74 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
75 vacc1x89ABCDEF = vminq_f16(vacc1x89ABCDEF, vmax);
76
77 vst1q_f16(o0, vacc0x01234567); o0 += 8;
78 vst1q_f16(o0, vacc0x89ABCDEF); o0 += 8;
79 vst1q_f16(o1, vacc1x01234567); o1 += 8;
80 vst1q_f16(o1, vacc1x89ABCDEF); o1 += 8;
81 }
82 for (; c >= 8 * sizeof(__fp16); c -= 8 * sizeof(__fp16)) {
83 const float16x8_t vscale01234567 = vld1q_f16(w);
84
85 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 += 8;
86 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 += 8;
87
88 const float16x8_t vbias01234567 = vld1q_f16(w + 16);
89 w += 8;
90
91 vacc0x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc0x01234567);
92 vacc1x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc1x01234567);
93
94 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
95 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
96
97 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
98 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
99
100 vst1q_f16(o0, vacc0x01234567); o0 += 8;
101 vst1q_f16(o1, vacc1x01234567); o1 += 8;
102 }
103 if XNN_UNLIKELY(c != 0) {
104 const float16x8_t vscale01234567 = vld1q_f16(w);
105
106 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 = (const __fp16*) ((uintptr_t) i0 + c);
107 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 = (const __fp16*) ((uintptr_t) i1 + c);
108
109 const float16x8_t vbias01234567 = vld1q_f16(w + 16);
110
111 vacc0x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc0x01234567);
112 vacc1x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc1x01234567);
113
114 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
115 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
116
117 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
118 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
119
120 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
121 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
122 if (c & (4 * sizeof(__fp16))) {
123 vst1_f16(o0, vacc0x0123); o0 += 4;
124 vst1_f16(o1, vacc1x0123); o1 += 4;
125
126 vacc0x0123 = vget_high_f16(vacc0x01234567);
127 vacc1x0123 = vget_high_f16(vacc1x01234567);
128 }
129 if (c & (2 * sizeof(__fp16))) {
130 vst1_lane_u32((void*) o0, vreinterpret_u32_f16(vacc0x0123), 0); o0 += 2;
131 vst1_lane_u32((void*) o1, vreinterpret_u32_f16(vacc1x0123), 0); o1 += 2;
132
133 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
134 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
135 }
136 if (c & (1 * sizeof(__fp16))) {
137 vst1_lane_f16(o0, vacc0x0123, 0); o0 += 1;
138 vst1_lane_f16(o1, vacc1x0123, 0); o1 += 1;
139 }
140 }
141 i0 = (const __fp16*) ((uintptr_t) i0 + input_increment);
142 o0 = (__fp16*) ((uintptr_t) o0 + output_increment);
143 i1 = (const __fp16*) ((uintptr_t) i1 + input_increment);
144 o1 = (__fp16*) ((uintptr_t) o1 + output_increment);
145 rows = doz(rows, 2);
146 } while (rows != 0);
147 }
148