xref: /aosp_15_r20/external/XNNPACK/src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vmulcaddc/neon.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 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/math.h>
15 #include <xnnpack/vmulcaddc.h>
16 
17 
xnn_f32_vmulcaddc_minmax_ukernel_c4__neonfma_2x(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vmulcaddc_minmax_ukernel_c4__neonfma_2x(
19     size_t rows,
20     size_t channels,
21     const float*restrict input,
22     size_t input_stride,
23     const float*restrict weights,
24     float*restrict output,
25     size_t output_stride,
26     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28   assert(rows != 0);
29   assert(channels != 0);
30   assert(channels % sizeof(float) == 0);
31 
32   const float* i0 = input;
33   float* o0 = output;
34   const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
35   float* o1 = (float*) ((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 float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
41   const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
42   do {
43     if XNN_UNPREDICTABLE(rows < 2) {
44       i1 = i0;
45       o1 = o0;
46     }
47 
48     const float* w = weights;
49     size_t c = channels;
50     for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
51       const float32x4_t vscale0123 = vld1q_f32(w); w += 4;
52 
53       float32x4_t vacc0x0123 = vld1q_f32(i0); i0 += 4;
54       float32x4_t vacc1x0123 = vld1q_f32(i1); i1 += 4;
55 
56 
57       const float32x4_t vbias0123 = vld1q_f32(w); w += 4;
58 
59       vacc0x0123 = vfmaq_f32(vbias0123, vscale0123, vacc0x0123);
60       vacc1x0123 = vfmaq_f32(vbias0123, vscale0123, vacc1x0123);
61 
62       vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
63       vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
64 
65       vacc0x0123 = vminq_f32(vacc0x0123, vmax);
66       vacc1x0123 = vminq_f32(vacc1x0123, vmax);
67 
68       vst1q_f32(o0, vacc0x0123); o0 += 4;
69       vst1q_f32(o1, vacc1x0123); o1 += 4;
70     }
71     if XNN_UNLIKELY(c != 0) {
72       const float32x4_t vscale0123 = vld1q_f32(w);
73 
74       float32x4_t vacc0x0123 = vld1q_f32(i0); i0 = (const float*) ((uintptr_t) i0 + c);
75       float32x4_t vacc1x0123 = vld1q_f32(i1); i1 = (const float*) ((uintptr_t) i1 + c);
76 
77 
78       const float32x4_t vbias0123 = vld1q_f32(w + 4);
79 
80       vacc0x0123 = vfmaq_f32(vbias0123, vscale0123, vacc0x0123);
81       vacc1x0123 = vfmaq_f32(vbias0123, vscale0123, vacc1x0123);
82 
83       vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
84       vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
85 
86       vacc0x0123 = vminq_f32(vacc0x0123, vmax);
87       vacc1x0123 = vminq_f32(vacc1x0123, vmax);
88 
89       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
90       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
91       if (c & (2 * sizeof(float))) {
92         vst1_f32(o0, vacc0x01); o0 += 2;
93         vst1_f32(o1, vacc1x01); o1 += 2;
94 
95         vacc0x01 = vget_high_f32(vacc0x0123);
96         vacc1x01 = vget_high_f32(vacc1x0123);
97       }
98       if (c & (1 * sizeof(float))) {
99         vst1_lane_f32(o0, vacc0x01, 0); o0 += 1;
100         vst1_lane_f32(o1, vacc1x01, 0); o1 += 1;
101       }
102     }
103     i0 = (const float*) ((uintptr_t) i0 + input_increment);
104     o0 = (float*) ((uintptr_t) o0 + output_increment);
105     i1 = (const float*) ((uintptr_t) i1 + input_increment);
106     o1 = (float*) ((uintptr_t) o1 + output_increment);
107     rows = doz(rows, 2);
108   } while (rows != 0);
109 }
110