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_x86_loadsplat_1x4_acc4(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_x86_loadsplat_1x4_acc4(
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)]) XNN_OOB_READS
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_v128_load32_splat(¶ms->scalar.max);
37 const v128_t vmin = wasm_v128_load32_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_v128_load64_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 size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
54
55 const float* i0 = zero;
56 const float* i1 = input;
57 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
58
59 float* o0 = output;
60
61 size_t output_height = input_height;
62 do {
63 if XNN_UNPREDICTABLE(output_height < 2) {
64 i2 = zero;
65 }
66
67 v128_t vi0x0123 = wasm_f32x4_const_splat(0.0f);
68 v128_t vi1x0123 = wasm_f32x4_const_splat(0.0f);
69 v128_t vi2x0123 = wasm_f32x4_const_splat(0.0f);
70
71 v128_t vi0x4567 = wasm_v128_load(i0);
72 i0 += 4;
73 v128_t vi1x4567 = wasm_v128_load(i1);
74 i1 += 4;
75 v128_t vi2x4567 = wasm_v128_load(i2);
76 i2 += 4;
77
78 size_t w = input_width;
79 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
80 const v128_t vi0x89AB = wasm_v128_load(i0);
81 i0 += 4;
82 const v128_t vi1x89AB = wasm_v128_load(i1);
83 i1 += 4;
84 const v128_t vi2x89AB = wasm_v128_load(i2);
85 i2 += 4;
86
87 v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
88 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, vk11);
89 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, vk21);
90
91 const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
92 const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
93 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
94
95 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, vk00);
96 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
97 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk20));
98
99 vi0x0123 = vi0x4567;
100 vi1x0123 = vi1x4567;
101 vi2x0123 = vi2x4567;
102
103 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
104 const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
105 const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
106
107 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk02));
108 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x5678, vk12));
109 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
110
111 vi0x4567 = vi0x89AB;
112 vi1x4567 = vi1x89AB;
113 vi2x4567 = vi2x89AB;
114
115 vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
116 vo0p2 = wasm_f32x4_add(vo0p2, vo0p3);
117 vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
118
119 v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
120 vo0 = wasm_f32x4_pmin(vmax, vo0);
121
122 wasm_v128_store(o0, vo0);
123 o0 += 4;
124 }
125 // Always process the last block of 1..4 pixels.
126 assert(w >= 1 * sizeof(float));
127 assert(w <= 4 * sizeof(float));
128 {
129 vi0x4567 = wasm_v128_and(vmask, vi0x4567);
130 vi1x4567 = wasm_v128_and(vmask, vi1x4567);
131 vi2x4567 = wasm_v128_and(vmask, vi2x4567);
132
133 v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
134 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, vk11);
135 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, vk21);
136
137 const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
138 const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
139 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
140
141 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, vk00);
142 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
143 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk20));
144
145 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
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 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk02));
151 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x5678, vk12));
152 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
153
154 vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
155 vo0p2 = wasm_f32x4_add(vo0p2, vo0p3);
156 vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
157
158 v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
159 vo0 = wasm_f32x4_pmin(vmax, vo0);
160
161 if XNN_LIKELY(w == 4 * sizeof(float)) {
162 wasm_v128_store(o0, vo0);
163 o0 += 4;
164 } else {
165 if (w & (2 * sizeof(float))) {
166 *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0);
167 o0 += 2;
168
169 vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
170 }
171 if (w & (1 * sizeof(float))) {
172 *o0 = wasm_f32x4_extract_lane(vo0, 0);
173 o0 += 1;
174 }
175 }
176 }
177
178 i0 = (const float*) ((uintptr_t) i1 - input_decrement);
179 i1 = (const float*) ((uintptr_t) i2 - input_decrement);
180 i2 = (const float*) ((uintptr_t) i1 + input_width);
181
182
183 } while (--output_height != 0);
184 }
185