1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/3x3p1-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_3x3p1__wasmsimd_x86_splat_3x4(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_3x3p1__wasmsimd_x86_splat_3x4(
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)])
28 {
29   assert(input_height != 0);
30   assert(input_width != 0);
31   assert(input_width % sizeof(float) == 0);
32   assert(padding_top == 1);
33 
34   const v128_t vmask = wasm_v128_load(params->scalar.mask);
35   const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
36   const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
37 
38   const v128_t vw0123 = wasm_v128_load(weights);
39   const v128_t vw4567 = wasm_v128_load(weights + 4);
40   const v128_t vw89 = wasm_v64x2_load_splat(weights + 8);
41 
42   const v128_t vzero = wasm_f32x4_splat(0.0f);
43 
44   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
45 
46   const float* i0 = zero;
47   const float* i1 = input;
48   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
49   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
50   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
51 
52   float* o0 = output;
53   float* o1 = (float*) ((uintptr_t) o0 + input_width);
54   float* o2 = (float*) ((uintptr_t) o1 + input_width);
55 
56   size_t output_height = input_height;
57   do {
58     if XNN_UNPREDICTABLE(output_height < 2) {
59       i2 = zero;
60       o1 = o0;
61     }
62     if XNN_UNPREDICTABLE(output_height < 3) {
63       i3 = zero;
64       o2 = o1;
65     }
66     if XNN_UNPREDICTABLE(output_height < 4) {
67       i4 = zero;
68     }
69 
70     v128_t vi0x0123 = vzero;
71     v128_t vi1x0123 = vzero;
72     v128_t vi2x0123 = vzero;
73     v128_t vi3x0123 = vzero;
74     v128_t vi4x0123 = vzero;
75 
76     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
77     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
78     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
79     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
80     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
81 
82     size_t w = input_width;
83     for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
84       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
85       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
86       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
87 
88       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
89       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
90       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
91       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
92       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
93 
94       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
95       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
96       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
97 
98       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
99       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
100       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
101 
102       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
103       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
104       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
105 
106       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
107       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
108       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
109       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
110       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
111 
112       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
113       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
114       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
115 
116       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
117       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
118       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
119 
120       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
121       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
122       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
123 
124       vi0x0123 = vi0x4567;
125       vi1x0123 = vi1x4567;
126       vi2x0123 = vi2x4567;
127       vi3x0123 = vi3x4567;
128       vi4x0123 = vi4x4567;
129 
130       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
131       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
132       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
133       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
134       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
135 
136       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
137       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
138       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
139 
140       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
141       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
142       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
143 
144       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
145       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
146       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
147 
148       vi0x4567 = vi0x89AB;
149       vi1x4567 = vi1x89AB;
150       vi2x4567 = vi2x89AB;
151       vi3x4567 = vi3x89AB;
152       vi4x4567 = vi4x89AB;
153 
154 
155       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
156       v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
157       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
158       vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
159       vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
160       vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));
161 
162       wasm_v128_store(o2, vo2); o2 += 4;
163       wasm_v128_store(o1, vo1); o1 += 4;
164       wasm_v128_store(o0, vo0); o0 += 4;
165     }
166     // Always process the last block of 1..4 pixels.
167     assert(w >= 1 * sizeof(float));
168     assert(w <= 4 * sizeof(float));
169     {
170       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
171       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
172       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
173 
174       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
175       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
176       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
177       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
178       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
179 
180       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
181       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
182       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
183 
184       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
185       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
186       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
187 
188       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
189       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
190       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
191 
192       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
193       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
194       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
195       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
196       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
197 
198       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
199       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
200       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
201 
202       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
203       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
204       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
205 
206       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
207       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
208       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
209 
210       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
211       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
212       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
213       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
214       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
215 
216       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
217       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
218       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
219 
220       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
221       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
222       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
223 
224       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
225       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
226       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
227 
228 
229       v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin));
230       v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin));
231       v128_t vo2 = wasm_v128_bitselect(vmin, vo2p0, wasm_f32x4_lt(vo2p0, vmin));
232       vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax));
233       vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax));
234       vo2 = wasm_v128_bitselect(vo2, vmax, wasm_f32x4_le(vo2, vmax));
235 
236       if XNN_LIKELY(w == 4 * sizeof(float)) {
237         wasm_v128_store(o2, vo2); o2 += 4;
238         wasm_v128_store(o1, vo1); o1 += 4;
239         wasm_v128_store(o0, vo0); o0 += 4;
240       } else {
241         if (w & (2 * sizeof(float))) {
242           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0); o2 += 2;
243           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
244           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
245 
246           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
247           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
248           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
249         }
250         if (w & (1 * sizeof(float))) {
251           *o2 = wasm_f32x4_extract_lane(vo2, 0); o2 += 1;
252           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
253           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
254         }
255       }
256     }
257 
258     i0 = (const float*) ((uintptr_t) i3 - input_decrement);
259     i1 = (const float*) ((uintptr_t) i4 - input_decrement);
260     i2 = (const float*) ((uintptr_t) i1 + input_width);
261     i3 = (const float*) ((uintptr_t) i2 + input_width);
262     i4 = (const float*) ((uintptr_t) i3 + input_width);
263 
264     o0 = o2;
265     o1 = (float*) ((uintptr_t) o0 + input_width);
266     o2 = (float*) ((uintptr_t) o1 + input_width);
267 
268     output_height = doz(output_height, 3);
269   } while (output_height != 0);
270 }
271