1 // Auto-generated file. Do not edit!
2 // Template: src/f32-bilinear/psimd.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 <psimd.h>
13
14 #include <xnnpack/bilinear.h>
15
16
xnn_f32_bilinear_ukernel__psimd_c8(size_t output_pixels,size_t channels,const float ** restrict input,size_t input_offset,const float * restrict weights,float * restrict output,size_t output_increment)17 void xnn_f32_bilinear_ukernel__psimd_c8(
18 size_t output_pixels,
19 size_t channels,
20 const float**restrict input,
21 size_t input_offset,
22 const float*restrict weights,
23 float*restrict output,
24 size_t output_increment)
25 {
26 assert(output_pixels != 0);
27 assert(channels != 0);
28 assert(channels % sizeof(float) == 0);
29
30 do {
31 const float* i0 = (const float*) ((uintptr_t) input[0] + input_offset);
32 const float* i1 = (const float*) ((uintptr_t) input[1] + input_offset);
33 const float* i2 = (const float*) ((uintptr_t) input[2] + input_offset);
34 const float* i3 = (const float*) ((uintptr_t) input[3] + input_offset);
35 input += 4;
36
37 const psimd_f32 valphah = psimd_load_splat_f32(weights);
38 const psimd_f32 valphav = psimd_load_splat_f32(weights + 1);
39 weights += 2;
40
41 size_t c = channels;
42 for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
43 const psimd_f32 vtl0123 = psimd_load_f32(i0);
44 const psimd_f32 vtr0123 = psimd_load_f32(i1);
45 const psimd_f32 vbl0123 = psimd_load_f32(i2);
46 const psimd_f32 vbr0123 = psimd_load_f32(i3);
47 const psimd_f32 vtl4567 = psimd_load_f32(i0 + 4);
48 const psimd_f32 vtr4567 = psimd_load_f32(i1 + 4);
49 const psimd_f32 vbl4567 = psimd_load_f32(i2 + 4);
50 const psimd_f32 vbr4567 = psimd_load_f32(i3 + 4);
51 i0 += 8;
52 i1 += 8;
53 i2 += 8;
54 i3 += 8;
55
56 const psimd_f32 vtd0123 = psimd_sub_f32(vtr0123, vtl0123);
57 const psimd_f32 vbd0123 = psimd_sub_f32(vbr0123, vbl0123);
58 const psimd_f32 vtd4567 = psimd_sub_f32(vtr4567, vtl4567);
59 const psimd_f32 vbd4567 = psimd_sub_f32(vbr4567, vbl4567);
60
61 const psimd_f32 vt0123 = psimd_qfma_f32(vtl0123, vtd0123, valphah);
62 const psimd_f32 vb0123 = psimd_qfma_f32(vbl0123, vbd0123, valphah);
63 const psimd_f32 vt4567 = psimd_qfma_f32(vtl4567, vtd4567, valphah);
64 const psimd_f32 vb4567 = psimd_qfma_f32(vbl4567, vbd4567, valphah);
65
66 const psimd_f32 vd0123 = psimd_sub_f32(vb0123, vt0123);
67 const psimd_f32 vd4567 = psimd_sub_f32(vb4567, vt4567);
68
69 const psimd_f32 vo0123 = psimd_qfma_f32(vt0123, vd0123, valphav);
70 const psimd_f32 vo4567 = psimd_qfma_f32(vt4567, vd4567, valphav);
71
72 psimd_store_f32(output, vo0123);
73 psimd_store_f32(output + 4, vo4567);
74 output += 8;
75 }
76 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
77 const psimd_f32 vtl0123 = psimd_load_f32(i0);
78 const psimd_f32 vtr0123 = psimd_load_f32(i1);
79 const psimd_f32 vbl0123 = psimd_load_f32(i2);
80 const psimd_f32 vbr0123 = psimd_load_f32(i3);
81 i0 += 4;
82 i1 += 4;
83 i2 += 4;
84 i3 += 4;
85
86 const psimd_f32 vtd0123 = psimd_sub_f32(vtr0123, vtl0123);
87 const psimd_f32 vbd0123 = psimd_sub_f32(vbr0123, vbl0123);
88
89 const psimd_f32 vt0123 = psimd_qfma_f32(vtl0123, vtd0123, valphah);
90 const psimd_f32 vb0123 = psimd_qfma_f32(vbl0123, vbd0123, valphah);
91
92 const psimd_f32 vd0123 = psimd_sub_f32(vb0123, vt0123);
93
94 const psimd_f32 vo0123 = psimd_qfma_f32(vt0123, vd0123, valphav);
95
96 psimd_store_f32(output, vo0123);
97 output += 4;
98 }
99 if XNN_UNLIKELY(c != 0) {
100 const psimd_f32 vtl0123 = psimd_load_f32(i0);
101 const psimd_f32 vtr0123 = psimd_load_f32(i1);
102 const psimd_f32 vbl0123 = psimd_load_f32(i2);
103 const psimd_f32 vbr0123 = psimd_load_f32(i3);
104
105 const psimd_f32 vtd0123 = psimd_sub_f32(vtr0123, vtl0123);
106 const psimd_f32 vbd0123 = psimd_sub_f32(vbr0123, vbl0123);
107
108 const psimd_f32 vt0123 = psimd_qfma_f32(vtl0123, vtd0123, valphah);
109 const psimd_f32 vb0123 = psimd_qfma_f32(vbl0123, vbd0123, valphah);
110
111 const psimd_f32 vd0123 = psimd_sub_f32(vb0123, vt0123);
112
113 psimd_f32 vo0123 = psimd_qfma_f32(vt0123, vd0123, valphav);
114
115 if (c & (2 * sizeof(float))) {
116 psimd_store2_f32(output, vo0123);
117 vo0123 = psimd_concat_hi_f32(vo0123, vo0123);
118 output += 2;
119 }
120 if (c & (1 * sizeof(float))) {
121 psimd_store1_f32(output, vo0123);
122 output += 1;
123 }
124 }
125
126 output = (float*) ((uintptr_t) output + output_increment);
127 } while (--output_pixels != 0);
128 }
129