1 // Auto-generated file. Do not edit!
2 // Template: src/f32-ibilinear/wasmsimd.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 <wasm_simd128.h>
13
14 #include <xnnpack/ibilinear.h>
15
16
xnn_f32_ibilinear_ukernel__wasmsimd_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_ibilinear_ukernel__wasmsimd_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) XNN_DISABLE_TSAN
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 v128_t valphah = wasm_v32x4_load_splat(weights);
38 const v128_t valphav = wasm_v32x4_load_splat(weights + 1);
39 weights += 2;
40
41 size_t c = channels;
42 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
43 const v128_t vtl = wasm_v128_load(i0);
44 const v128_t vtr = wasm_v128_load(i1);
45 const v128_t vbl = wasm_v128_load(i2);
46 const v128_t vbr = wasm_v128_load(i3);
47 i0 += 4;
48 i1 += 4;
49 i2 += 4;
50 i3 += 4;
51
52 const v128_t vtd = wasm_f32x4_sub(vtr, vtl);
53 const v128_t vbd = wasm_f32x4_sub(vbr, vbl);
54 const v128_t vt = wasm_f32x4_add(vtl, wasm_f32x4_mul(vtd, valphah));
55 const v128_t vb = wasm_f32x4_add(vbl, wasm_f32x4_mul(vbd, valphah));
56 const v128_t vd = wasm_f32x4_sub(vb, vt);
57 const v128_t vo = wasm_f32x4_add(vt, wasm_f32x4_mul(vd, valphav));
58
59 wasm_v128_store(output, vo);
60 output += 4;
61 }
62 if XNN_UNLIKELY(c != 0) {
63 const v128_t vtl = wasm_v128_load(i0);
64 const v128_t vtr = wasm_v128_load(i1);
65 const v128_t vbl = wasm_v128_load(i2);
66 const v128_t vbr = wasm_v128_load(i3);
67
68 const v128_t vtd = wasm_f32x4_sub(vtr, vtl);
69 const v128_t vbd = wasm_f32x4_sub(vbr, vbl);
70 const v128_t vt = wasm_f32x4_add(vtl, wasm_f32x4_mul(vtd, valphah));
71 const v128_t vb = wasm_f32x4_add(vbl, wasm_f32x4_mul(vbd, valphah));
72 const v128_t vd = wasm_f32x4_sub(vb, vt);
73 v128_t vo = wasm_f32x4_add(vt, wasm_f32x4_mul(vd, valphav));
74
75 if (c & (2 * sizeof(float))) {
76 *((double*) output) = wasm_f64x2_extract_lane(vo, 0);
77 vo = wasm_v32x4_shuffle(vo, vo, 2, 3, 2, 3);
78 output += 2;
79 }
80 if (c & (1 * sizeof(float))) {
81 *output++ = wasm_f32x4_extract_lane(vo, 0);
82 }
83 }
84
85 output = (float*) ((uintptr_t) output + output_increment);
86 } while (--output_pixels != 0);
87 }
88