1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-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/dwconv.h>
15
16
xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm(
18 size_t channels,
19 size_t output_width,
20 const float** input,
21 const float* weights,
22 float* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const float* zero,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min);
33 const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.max);
34 do {
35 const float* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const float*) ((uintptr_t) i0 + input_offset);
39 }
40 const float* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 }
45 const float* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const float*) ((uintptr_t) i2 + input_offset);
49 }
50 const float* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const float*) ((uintptr_t) i3 + input_offset);
54 }
55 input = (const float**) ((uintptr_t) input + input_stride);
56
57 size_t c = channels;
58 const float* w = weights;
59 for (; c >= 4; c -= 4) {
60 v128_t vacc0123p0 = wasm_v128_load(w);
61
62
63 const v128_t vi0x0123 = wasm_v128_load(i0);
64 i0 += 4;
65
66 const v128_t vk0x0123 = wasm_v128_load(w + 4);
67 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
68
69 const v128_t vi1x0123 = wasm_v128_load(i1);
70 i1 += 4;
71
72 const v128_t vk1x0123 = wasm_v128_load(w + 8);
73 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
74
75 const v128_t vi2x0123 = wasm_v128_load(i2);
76 i2 += 4;
77
78 const v128_t vk2x0123 = wasm_v128_load(w + 12);
79 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
80
81 const v128_t vi3x0123 = wasm_v128_load(i3);
82 i3 += 4;
83
84 const v128_t vk3x0123 = wasm_v128_load(w + 16);
85 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
86
87 w += 20;
88
89
90 v128_t vacc0123 = wasm_f32x4_max(vacc0123p0, vmin);
91
92 vacc0123 = wasm_f32x4_min(vacc0123, vmax);
93
94 wasm_v128_store(output, vacc0123);
95 output += 4;
96 }
97 if XNN_UNLIKELY(c != 0) {
98 v128_t vacc0123p0 = wasm_v128_load(w);
99
100 const v128_t vi0x0123 = wasm_v128_load(i0);
101 const v128_t vk0x0123 = wasm_v128_load(w + 4);
102 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
103
104 const v128_t vi1x0123 = wasm_v128_load(i1);
105 const v128_t vk1x0123 = wasm_v128_load(w + 8);
106 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
107
108 const v128_t vi2x0123 = wasm_v128_load(i2);
109 const v128_t vk2x0123 = wasm_v128_load(w + 12);
110 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
111
112 const v128_t vi3x0123 = wasm_v128_load(i3);
113 const v128_t vk3x0123 = wasm_v128_load(w + 16);
114 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
115
116
117 v128_t vacc0123 = wasm_f32x4_max(vacc0123p0, vmin);
118 vacc0123 = wasm_f32x4_min(vacc0123, vmax);
119
120 if (c & 2) {
121 *((double*) output) = wasm_f64x2_extract_lane(vacc0123, 0);
122 vacc0123 = wasm_v32x4_shuffle(vacc0123, vacc0123, 2, 3, 2, 3);
123 output += 2;
124 }
125 if (c & 1) {
126 *output = wasm_f32x4_extract_lane(vacc0123, 0);
127 output += 1;
128 }
129 }
130
131 output = (float*) ((uintptr_t) output + output_increment);
132 } while (--output_width != 0);
133 }
134