xref: /aosp_15_r20/external/XNNPACK/src/f32-dwconv/gen/up4x4-wasmsimd.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_ukernel_up4x4__wasmsimd(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_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_ukernel_up4x4__wasmsimd(
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_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(channels != 0);
30   assert(output_width != 0);
31 
32   do {
33     const float* i0 = input[0];
34     assert(i0 != NULL);
35     if XNN_UNPREDICTABLE(i0 != zero) {
36       i0 = (const float*) ((uintptr_t) i0 + input_offset);
37     }
38     const float* i1 = input[1];
39     assert(i1 != NULL);
40     if XNN_UNPREDICTABLE(i1 != zero) {
41       i1 = (const float*) ((uintptr_t) i1 + input_offset);
42     }
43     const float* i2 = input[2];
44     assert(i2 != NULL);
45     if XNN_UNPREDICTABLE(i2 != zero) {
46       i2 = (const float*) ((uintptr_t) i2 + input_offset);
47     }
48     const float* i3 = input[3];
49     assert(i3 != NULL);
50     if XNN_UNPREDICTABLE(i3 != zero) {
51       i3 = (const float*) ((uintptr_t) i3 + input_offset);
52     }
53     input = (const float**) ((uintptr_t) input + input_stride);
54 
55     size_t c = channels;
56     const float* w = weights;
57     for (; c >= 4; c -= 4) {
58       v128_t vacc0123p0 = wasm_v128_load(w);
59 
60 
61       const v128_t vi0x0123 = wasm_v128_load(i0);
62       i0 += 4;
63 
64       const v128_t vk0x0123 = wasm_v128_load(w + 4);
65       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
66 
67       const v128_t vi1x0123 = wasm_v128_load(i1);
68       i1 += 4;
69 
70       const v128_t vk1x0123 = wasm_v128_load(w + 8);
71       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
72 
73       const v128_t vi2x0123 = wasm_v128_load(i2);
74       i2 += 4;
75 
76       const v128_t vk2x0123 = wasm_v128_load(w + 12);
77       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
78 
79       const v128_t vi3x0123 = wasm_v128_load(i3);
80       i3 += 4;
81 
82       const v128_t vk3x0123 = wasm_v128_load(w + 16);
83       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
84 
85       w += 20;
86 
87 
88       const v128_t vacc0123 = vacc0123p0;
89 
90       wasm_v128_store(output, vacc0123);
91       output += 4;
92     }
93     if XNN_UNLIKELY(c != 0) {
94       v128_t vacc0123p0 = wasm_v128_load(w);
95 
96       const v128_t vi0x0123 = wasm_v128_load(i0);
97       const v128_t vk0x0123 = wasm_v128_load(w + 4);
98       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi0x0123, vk0x0123));
99 
100       const v128_t vi1x0123 = wasm_v128_load(i1);
101       const v128_t vk1x0123 = wasm_v128_load(w + 8);
102       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123));
103 
104       const v128_t vi2x0123 = wasm_v128_load(i2);
105       const v128_t vk2x0123 = wasm_v128_load(w + 12);
106       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi2x0123, vk2x0123));
107 
108       const v128_t vi3x0123 = wasm_v128_load(i3);
109       const v128_t vk3x0123 = wasm_v128_load(w + 16);
110       vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi3x0123, vk3x0123));
111 
112 
113       v128_t vacc0123 = vacc0123p0;
114 
115       if (c & 2) {
116         *((double*) output) = wasm_f64x2_extract_lane(vacc0123, 0);
117         vacc0123 = wasm_v32x4_shuffle(vacc0123, vacc0123, 2, 3, 2, 3);
118         output += 2;
119       }
120       if (c & 1) {
121         *output = wasm_f32x4_extract_lane(vacc0123, 0);
122         output += 1;
123       }
124     }
125 
126     output = (float*) ((uintptr_t) output + output_increment);
127   } while (--output_width != 0);
128 }
129