1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/3x3p1-wasmsimd-loadsplat.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
15 #include <xnnpack/dwconv.h>
16 #include <xnnpack/math.h>
17
18
19
xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3(
21 size_t input_height,
22 size_t input_width,
23 const float* input,
24 const float* weights,
25 const float* zero,
26 float* output,
27 uint32_t padding_top,
28 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(input_height != 0);
31 assert(input_width != 0);
32 assert(input_width % sizeof(float) == 0);
33 assert(padding_top == 1);
34
35 const v128_t vmask = wasm_v128_load(params->scalar.mask);
36 const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.max);
37 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min);
38
39 const v128_t vw0123 = wasm_v128_load(weights);
40 const v128_t vw4567 = wasm_v128_load(weights + 4);
41 const v128_t vw89 = wasm_v64x2_load_splat(weights + 8);
42 const v128_t vbias = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
43 const v128_t vk00 = wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1);
44 const v128_t vk01 = wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2);
45 const v128_t vk02 = wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3);
46 const v128_t vk10 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0);
47 const v128_t vk11 = wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1);
48 const v128_t vk12 = wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2);
49 const v128_t vk20 = wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3);
50 const v128_t vk21 = wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0);
51 const v128_t vk22 = wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1);
52
53 const v128_t vzero = wasm_f32x4_splat(0.0f);
54
55 const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
56
57 const float* i0 = zero;
58 const float* i1 = input;
59 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
60
61 float* o0 = output;
62
63 size_t output_height = input_height;
64 do {
65 if XNN_UNPREDICTABLE(output_height < 2) {
66 i2 = zero;
67 }
68
69 v128_t vi0x0123 = vzero;
70 v128_t vi1x0123 = vzero;
71 v128_t vi2x0123 = vzero;
72
73 v128_t vi0x4567 = wasm_v128_load(i0);
74 i0 += 4;
75 v128_t vi1x4567 = wasm_v128_load(i1);
76 i1 += 4;
77 v128_t vi2x4567 = wasm_v128_load(i2);
78 i2 += 4;
79
80 size_t w = input_width;
81 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
82 const v128_t vi0x89AB = wasm_v128_load(i0);
83 i0 += 4;
84 const v128_t vi1x89AB = wasm_v128_load(i1);
85 i1 += 4;
86 const v128_t vi2x89AB = wasm_v128_load(i2);
87 i2 += 4;
88
89 v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
90 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, vk11);
91 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, vk21);
92
93 const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
94 const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
95 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
96
97 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00));
98 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x3456, vk10));
99 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x3456, vk20));
100
101 vi0x0123 = vi0x4567;
102 vi1x0123 = vi1x4567;
103 vi2x0123 = vi2x4567;
104
105 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
106 const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
107 const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
108
109 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
110 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x5678, vk12));
111 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x5678, vk22));
112
113 vi0x4567 = vi0x89AB;
114 vi1x4567 = vi1x89AB;
115 vi2x4567 = vi2x89AB;
116
117 vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
118 vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
119
120 v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
121 vo0 = wasm_f32x4_min(vo0, vmax);
122
123 wasm_v128_store(o0, vo0);
124 o0 += 4;
125 }
126 // Always process the last block of 1..4 pixels.
127 assert(w >= 1 * sizeof(float));
128 assert(w <= 4 * sizeof(float));
129 {
130 vi0x4567 = wasm_v128_and(vmask, vi0x4567);
131 vi1x4567 = wasm_v128_and(vmask, vi1x4567);
132 vi2x4567 = wasm_v128_and(vmask, vi2x4567);
133
134 v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
135 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, vk11);
136 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, vk21);
137
138 const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
139 const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
140 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
141
142 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00));
143 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x3456, vk10));
144 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x3456, vk20));
145
146 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
147 const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
148 const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
149
150 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
151 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x5678, vk12));
152 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x5678, vk22));
153
154 vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
155 vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
156
157 v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
158 vo0 = wasm_f32x4_min(vo0, vmax);
159
160 if XNN_LIKELY(w == 4 * sizeof(float)) {
161 wasm_v128_store(o0, vo0);
162 o0 += 4;
163 } else {
164 if (w & (2 * sizeof(float))) {
165 *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0);
166 o0 += 2;
167
168 vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
169 }
170 if (w & (1 * sizeof(float))) {
171 *o0 = wasm_f32x4_extract_lane(vo0, 0);
172 o0 += 1;
173 }
174 }
175 }
176
177 i0 = (const float*) ((uintptr_t) i1 - input_decrement);
178 i1 = (const float*) ((uintptr_t) i2 - input_decrement);
179 i2 = (const float*) ((uintptr_t) i1 + input_width);
180
181
182 } while (--output_height != 0);
183 }
184