1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/3x3s2p1-wasmsimd-splat.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 #include <xnnpack/math.h>
16
17
18
xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2(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)])19 void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2(
20 size_t input_height,
21 size_t input_width,
22 const float* input,
23 const float* weights,
24 const float* zero,
25 float* output,
26 uint32_t padding_top,
27 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(input_height != 0);
30 assert(input_width != 0);
31 assert(input_width % sizeof(float) == 0);
32 assert(padding_top >= 0);
33 assert(padding_top <= 1);
34
35 const v128_t vmask_even = wasm_v128_load(params->scalar.mask_even);
36 const v128_t vmask_odd = wasm_v128_load(params->scalar.mask_odd);
37 const v128_t vmax = wasm_v128_load32_splat(¶ms->scalar.max);
38 const v128_t vmin = wasm_v128_load32_splat(¶ms->scalar.min);
39
40 const v128_t vw0123 = wasm_v128_load(weights);
41 const v128_t vw4567 = wasm_v128_load(weights + 4);
42 const v128_t vw89 = wasm_v128_load64_splat(weights + 8);
43
44 const size_t input_decrement = round_down_po2(input_width, 4 /* SIMD output width */ * 2 /* subsampling */ * sizeof(float));
45
46 const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
47 const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
48 if XNN_UNPREDICTABLE(padding_top != 0) {
49 i0 = zero;
50 }
51 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
52
53 float* o0 = output;
54
55 size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
56 size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
57 do {
58 if XNN_UNPREDICTABLE(padded_input_height < 4) {
59 i2 = zero;
60 }
61
62 v128_t vi0x1357 = wasm_f32x4_const_splat(0.0f);
63 v128_t vi1x1357 = wasm_f32x4_const_splat(0.0f);
64 v128_t vi2x1357 = wasm_f32x4_const_splat(0.0f);
65
66 size_t w = input_width;
67 for (; w >= 8 * sizeof(float); w -= 8 * sizeof(float)) {
68 v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
69
70 const v128_t vi0x89AB = wasm_v128_load(i0);
71 const v128_t vi0xCDEF = wasm_v128_load(i0 + 4);
72 i0 += 8;
73 const v128_t vi1x89AB = wasm_v128_load(i1);
74 const v128_t vi1xCDEF = wasm_v128_load(i1 + 4);
75 i1 += 8;
76 const v128_t vi2x89AB = wasm_v128_load(i2);
77 const v128_t vi2xCDEF = wasm_v128_load(i2 + 4);
78 i2 += 8;
79
80 const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6);
81 const v128_t vi0x9BDF = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 1, 3, 5, 7);
82 const v128_t vi1x8ACE = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 0, 2, 4, 6);
83 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7);
84 const v128_t vi2x8ACE = wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 0, 2, 4, 6);
85 const v128_t vi2x9BDF = wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 1, 3, 5, 7);
86
87 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2));
88
89 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
90
91 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
92
93 const v128_t vi0x7BDF = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6);
94 vi0x1357 = vi0x9BDF;
95 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6);
96 vi1x1357 = vi1x9BDF;
97 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6);
98 vi2x1357 = vi2x9BDF;
99
100 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
101
102 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
103
104 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
105
106 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x9BDF, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
107
108 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
109
110 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x9BDF, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
111
112 vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
113
114 v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
115 vo0 = wasm_f32x4_pmin(vmax, vo0);
116
117 wasm_v128_store(o0, vo0); o0 += 4;
118 }
119 // Last block has 0-7 pixels to process.
120 assert(w < 8 * sizeof(float));
121 if XNN_LIKELY(w != 0) {
122 v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
123
124 const v128_t vi0x89AB = wasm_v128_load(i0);
125 const v128_t vi0xCDEF = wasm_v128_load(i0 + 4);
126 const v128_t vi1x89AB = wasm_v128_load(i1);
127 const v128_t vi1xCDEF = wasm_v128_load(i1 + 4);
128 const v128_t vi2x89AB = wasm_v128_load(i2);
129 const v128_t vi2xCDEF = wasm_v128_load(i2 + 4);
130
131 const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6));
132 const v128_t vi0x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 1, 3, 5, 7));
133 const v128_t vi1x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 0, 2, 4, 6));
134 const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7));
135 const v128_t vi2x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 0, 2, 4, 6));
136 const v128_t vi2x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 1, 3, 5, 7));
137
138 v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2));
139
140 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
141
142 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
143
144 const v128_t vi0x7BDF = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6);
145 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6);
146 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6);
147
148 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
149
150 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
151
152 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
153
154 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x9BDF, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
155
156 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
157
158 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x9BDF, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
159
160 vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
161
162 v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
163 vo0 = wasm_f32x4_pmin(vmax, vo0);
164
165 w += 1 * sizeof(float);
166 if (w & (8 * sizeof(float))) {
167 wasm_v128_store(o0, vo0); o0 += 4;
168 } else {
169 if (w & (4 * sizeof(float))) {
170 *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
171
172 vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
173 }
174 if (w & (2 * sizeof(float))) {
175 *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
176 }
177 }
178 }
179
180 i0 = (const float*) ((uintptr_t) i2 - input_decrement);
181 i1 = (const float*) ((uintptr_t) i0 + input_width);
182 i2 = (const float*) ((uintptr_t) i1 + input_width);
183
184
185 output_height -= 1;
186 padded_input_height -= 2;
187 } while (output_height != 0);
188 }
189