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 struct 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 struct xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
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 if XNN_UNPREDICTABLE(rows < 2) {
37 i1 = i0;
38 o1 = o0;
39 }
40
41 const size_t input_increment = input_stride * 2 - channels;
42 const size_t output_increment = output_stride * 2 - channels;
43
44 const float16x8_t vmin = vld1q_dup_f16(¶ms->min);
45 const float16x8_t vmax = vld1q_dup_f16(¶ms->max);
46 do {
47 const __fp16* w = (const __fp16*) weights;
48 size_t c = channels;
49 for (; c >= 16 * sizeof(__fp16); c -= 16 * sizeof(__fp16)) {
50 const float16x8_t vscale01234567 = vld1q_f16(w); w += 8;
51 const float16x8_t vscale89ABCDEF = vld1q_f16(w); w += 8;
52
53 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 += 8;
54 float16x8_t vacc0x89ABCDEF = vld1q_f16(i0); i0 += 8;
55 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 += 8;
56 float16x8_t vacc1x89ABCDEF = vld1q_f16(i1); i1 += 8;
57
58 const float16x8_t vbias01234567 = vld1q_f16(w); w += 8;
59 const float16x8_t vbias89ABCDEF = vld1q_f16(w); w += 8;
60
61 vacc0x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc0x01234567);
62 vacc0x89ABCDEF = vfmaq_f16(vbias89ABCDEF, vscale89ABCDEF, vacc0x89ABCDEF);
63 vacc1x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc1x01234567);
64 vacc1x89ABCDEF = vfmaq_f16(vbias89ABCDEF, vscale89ABCDEF, vacc1x89ABCDEF);
65
66 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
67 vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
68 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
69 vacc1x89ABCDEF = vmaxq_f16(vacc1x89ABCDEF, vmin);
70
71 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
72 vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
73 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
74 vacc1x89ABCDEF = vminq_f16(vacc1x89ABCDEF, vmax);
75
76 vst1q_f16(o0, vacc0x01234567); o0 += 8;
77 vst1q_f16(o0, vacc0x89ABCDEF); o0 += 8;
78 vst1q_f16(o1, vacc1x01234567); o1 += 8;
79 vst1q_f16(o1, vacc1x89ABCDEF); o1 += 8;
80 }
81 for (; c >= 8 * sizeof(__fp16); c -= 8 * sizeof(__fp16)) {
82 const float16x8_t vscale01234567 = vld1q_f16(w); w += 8;
83
84 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 += 8;
85 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 += 8;
86
87 const float16x8_t vbias01234567 = vld1q_f16(w + 8);
88
89 vacc0x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc0x01234567);
90 vacc1x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc1x01234567);
91
92 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
93 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
94
95 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
96 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
97
98 vst1q_f16(o0, vacc0x01234567); o0 += 8;
99 vst1q_f16(o1, vacc1x01234567); o1 += 8;
100 }
101 if XNN_UNLIKELY(c != 0) {
102 const float16x8_t vscale01234567 = vld1q_f16(w);
103
104 float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 = (const __fp16*) ((uintptr_t) i0 + c);
105 float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 = (const __fp16*) ((uintptr_t) i1 + c);
106
107 const float16x8_t vbias01234567 = vld1q_f16(w + 16);
108
109 vacc0x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc0x01234567);
110 vacc1x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc1x01234567);
111
112 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
113 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
114
115 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
116 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
117
118 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
119 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
120 if (c & (4 * sizeof(__fp16))) {
121 vst1_f16(o0, vacc0x0123); o0 += 4;
122 vst1_f16(o1, vacc1x0123); o1 += 4;
123
124 vacc0x0123 = vget_high_f16(vacc0x01234567);
125 vacc1x0123 = vget_high_f16(vacc1x01234567);
126 }
127 if (c & (2 * sizeof(__fp16))) {
128 vst1_lane_u32(__builtin_assume_aligned(o0, 1), vreinterpret_u32_f16(vacc0x0123), 0); o0 += 2;
129 vst1_lane_u32(__builtin_assume_aligned(o1, 1), vreinterpret_u32_f16(vacc1x0123), 0); o1 += 2;
130
131 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
132 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
133 }
134 if (c & (1 * sizeof(__fp16))) {
135 vst1_lane_f16(o0, vacc0x0123, 0); o0 += 1;
136 vst1_lane_f16(o1, vacc1x0123, 0); o1 += 1;
137 }
138 }
139 i0 = (const __fp16*) ((uintptr_t) i0 + input_increment);
140 o0 = (__fp16*) ((uintptr_t) o0 + output_increment);
141 i1 = (const __fp16*) ((uintptr_t) i1 + input_increment);
142 o1 = (__fp16*) ((uintptr_t) o1 + output_increment);
143 if XNN_UNPREDICTABLE(rows < 4) {
144 i1 = i0;
145 o1 = o0;
146 }
147 rows = doz(rows, 2);
148 } while (rows != 0);
149 }
150