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_c4(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_c4(
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 >= 4 * sizeof(float); c -= 4 * 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 i0 += 4;
48 i1 += 4;
49 i2 += 4;
50 i3 += 4;
51
52 const psimd_f32 vtd0123 = psimd_sub_f32(vtr0123, vtl0123);
53 const psimd_f32 vbd0123 = psimd_sub_f32(vbr0123, vbl0123);
54
55 const psimd_f32 vt0123 = psimd_qfma_f32(vtl0123, vtd0123, valphah);
56 const psimd_f32 vb0123 = psimd_qfma_f32(vbl0123, vbd0123, valphah);
57
58 const psimd_f32 vd0123 = psimd_sub_f32(vb0123, vt0123);
59
60 const psimd_f32 vo0123 = psimd_qfma_f32(vt0123, vd0123, valphav);
61
62 psimd_store_f32(output, vo0123);
63 output += 4;
64 }
65 if XNN_UNLIKELY(c != 0) {
66 const psimd_f32 vtl0123 = psimd_load_f32(i0);
67 const psimd_f32 vtr0123 = psimd_load_f32(i1);
68 const psimd_f32 vbl0123 = psimd_load_f32(i2);
69 const psimd_f32 vbr0123 = psimd_load_f32(i3);
70
71 const psimd_f32 vtd0123 = psimd_sub_f32(vtr0123, vtl0123);
72 const psimd_f32 vbd0123 = psimd_sub_f32(vbr0123, vbl0123);
73
74 const psimd_f32 vt0123 = psimd_qfma_f32(vtl0123, vtd0123, valphah);
75 const psimd_f32 vb0123 = psimd_qfma_f32(vbl0123, vbd0123, valphah);
76
77 const psimd_f32 vd0123 = psimd_sub_f32(vb0123, vt0123);
78
79 psimd_f32 vo0123 = psimd_qfma_f32(vt0123, vd0123, valphav);
80
81 if (c & (2 * sizeof(float))) {
82 psimd_store2_f32(output, vo0123);
83 vo0123 = psimd_concat_hi_f32(vo0123, vo0123);
84 output += 2;
85 }
86 if (c & (1 * sizeof(float))) {
87 psimd_store1_f32(output, vo0123);
88 output += 1;
89 }
90 }
91
92 output = (float*) ((uintptr_t) output + output_increment);
93 } while (--output_pixels != 0);
94 }
95