xref: /aosp_15_r20/external/XNNPACK/src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_c8__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_c8__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(&params->neon.min));
41   const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->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 >= 8 * sizeof(__fp16); c -= 8 * sizeof(__fp16)) {
51       const float16x8_t vscale01234567 = vld1q_f16(w);
52 
53       float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 += 8;
54       float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 += 8;
55 
56       const float16x8_t vbias01234567 = vld1q_f16(w + 8);
57       w += 16;
58 
59       vacc0x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc0x01234567);
60       vacc1x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc1x01234567);
61 
62       vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
63       vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
64 
65       vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
66       vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
67 
68       vst1q_f16(o0, vacc0x01234567); o0 += 8;
69       vst1q_f16(o1, vacc1x01234567); o1 += 8;
70     }
71     if XNN_UNLIKELY(c != 0) {
72       const float16x8_t vscale01234567 = vld1q_f16(w);
73 
74       float16x8_t vacc0x01234567 = vld1q_f16(i0); i0 = (const __fp16*) ((uintptr_t) i0 + c);
75       float16x8_t vacc1x01234567 = vld1q_f16(i1); i1 = (const __fp16*) ((uintptr_t) i1 + c);
76 
77       const float16x8_t vbias01234567 = vld1q_f16(w + 8);
78 
79       vacc0x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc0x01234567);
80       vacc1x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc1x01234567);
81 
82       vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
83       vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
84 
85       vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
86       vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
87 
88       float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
89       float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
90       if (c & (4 * sizeof(__fp16))) {
91         vst1_f16(o0, vacc0x0123); o0 += 4;
92         vst1_f16(o1, vacc1x0123); o1 += 4;
93 
94         vacc0x0123 = vget_high_f16(vacc0x01234567);
95         vacc1x0123 = vget_high_f16(vacc1x01234567);
96       }
97       if (c & (2 * sizeof(__fp16))) {
98         vst1_lane_u32((void*) o0, vreinterpret_u32_f16(vacc0x0123), 0); o0 += 2;
99         vst1_lane_u32((void*) o1, vreinterpret_u32_f16(vacc1x0123), 0); o1 += 2;
100 
101         vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
102         vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
103       }
104       if (c & (1 * sizeof(__fp16))) {
105         vst1_lane_f16(o0, vacc0x0123, 0); o0 += 1;
106         vst1_lane_f16(o1, vacc1x0123, 0); o1 += 1;
107       }
108     }
109     i0 = (const __fp16*) ((uintptr_t) i0 + input_increment);
110     o0 = (__fp16*) ((uintptr_t) o0 + output_increment);
111     i1 = (const __fp16*) ((uintptr_t) i1 + input_increment);
112     o1 = (__fp16*) ((uintptr_t) o1 + output_increment);
113     rows = doz(rows, 2);
114   } while (rows != 0);
115 }
116