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_c8__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_c8__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(¶ms->scalar.min);
41 const float32x4_t vmax = vld1q_dup_f32(¶ms->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 >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
51 const float32x4_t vscale0123 = vld1q_f32(w); w += 4;
52 const float32x4_t vscale4567 = vld1q_f32(w); w += 4;
53
54 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 += 4;
55 float32x4_t vacc0x4567 = vld1q_f32(i0); i0 += 4;
56 float32x4_t vacc1x0123 = vld1q_f32(i1); i1 += 4;
57 float32x4_t vacc1x4567 = vld1q_f32(i1); i1 += 4;
58
59
60 const float32x4_t vbias0123 = vld1q_f32(w); w += 4;
61 const float32x4_t vbias4567 = vld1q_f32(w); w += 4;
62
63 vacc0x0123 = vfmaq_f32(vbias0123, vscale0123, vacc0x0123);
64 vacc0x4567 = vfmaq_f32(vbias4567, vscale4567, vacc0x4567);
65 vacc1x0123 = vfmaq_f32(vbias0123, vscale0123, vacc1x0123);
66 vacc1x4567 = vfmaq_f32(vbias4567, vscale4567, vacc1x4567);
67
68 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
69 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
70 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
71 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
72
73 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
74 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
75 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
76 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
77
78 vst1q_f32(o0, vacc0x0123); o0 += 4;
79 vst1q_f32(o0, vacc0x4567); o0 += 4;
80 vst1q_f32(o1, vacc1x0123); o1 += 4;
81 vst1q_f32(o1, vacc1x4567); o1 += 4;
82 }
83 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
84 const float32x4_t vscale0123 = vld1q_f32(w); w += 4;
85
86 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 += 4;
87 float32x4_t vacc1x0123 = vld1q_f32(i1); i1 += 4;
88
89
90 const float32x4_t vbias0123 = vld1q_f32(w + 4);
91
92 vacc0x0123 = vfmaq_f32(vbias0123, vscale0123, vacc0x0123);
93 vacc1x0123 = vfmaq_f32(vbias0123, vscale0123, vacc1x0123);
94
95 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
96 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
97
98 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
99 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
100
101 vst1q_f32(o0, vacc0x0123); o0 += 4;
102 vst1q_f32(o1, vacc1x0123); o1 += 4;
103 }
104 if XNN_UNLIKELY(c != 0) {
105 const float32x4_t vscale0123 = vld1q_f32(w);
106
107 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 = (const float*) ((uintptr_t) i0 + c);
108 float32x4_t vacc1x0123 = vld1q_f32(i1); i1 = (const float*) ((uintptr_t) i1 + c);
109
110
111 const float32x4_t vbias0123 = vld1q_f32(w + 8);
112
113 vacc0x0123 = vfmaq_f32(vbias0123, vscale0123, vacc0x0123);
114 vacc1x0123 = vfmaq_f32(vbias0123, vscale0123, vacc1x0123);
115
116 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
117 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
118
119 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
120 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
121
122 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
123 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
124 if (c & (2 * sizeof(float))) {
125 vst1_f32(o0, vacc0x01); o0 += 2;
126 vst1_f32(o1, vacc1x01); o1 += 2;
127
128 vacc0x01 = vget_high_f32(vacc0x0123);
129 vacc1x01 = vget_high_f32(vacc1x0123);
130 }
131 if (c & (1 * sizeof(float))) {
132 vst1_lane_f32(o0, vacc0x01, 0); o0 += 1;
133 vst1_lane_f32(o1, vacc1x01, 0); o1 += 1;
134 }
135 }
136 i0 = (const float*) ((uintptr_t) i0 + input_increment);
137 o0 = (float*) ((uintptr_t) o0 + output_increment);
138 i1 = (const float*) ((uintptr_t) i1 + input_increment);
139 o1 = (float*) ((uintptr_t) o1 + output_increment);
140 rows = doz(rows, 2);
141 } while (rows != 0);
142 }
143