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_2x4_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)])20 void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2(
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(&params->scalar.max);
37   const v128_t vmin = wasm_v128_load32_splat(&params->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   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
59 
60   float* o0 = output;
61   float* o1 = (float*) ((uintptr_t) o0 + input_width);
62 
63   size_t output_height = input_height;
64   do {
65     if XNN_UNPREDICTABLE(output_height < 2) {
66       i2 = zero;
67       o1 = o0;
68     }
69     if XNN_UNPREDICTABLE(output_height < 3) {
70       i3 = zero;
71     }
72 
73     v128_t vi0x0123 = wasm_f32x4_const_splat(0.0f);
74     v128_t vi1x0123 = wasm_f32x4_const_splat(0.0f);
75     v128_t vi2x0123 = wasm_f32x4_const_splat(0.0f);
76     v128_t vi3x0123 = wasm_f32x4_const_splat(0.0f);
77 
78     v128_t vi0x4567 = wasm_v128_load(i0);
79     i0 += 4;
80     v128_t vi1x4567 = wasm_v128_load(i1);
81     i1 += 4;
82     v128_t vi2x4567 = wasm_v128_load(i2);
83     i2 += 4;
84     v128_t vi3x4567 = wasm_v128_load(i3);
85     i3 += 4;
86 
87     size_t w = input_width;
88     for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
89       const v128_t vi0x89AB = wasm_v128_load(i0);
90       i0 += 4;
91       const v128_t vi1x89AB = wasm_v128_load(i1);
92       i1 += 4;
93       const v128_t vi2x89AB = wasm_v128_load(i2);
94       i2 += 4;
95       const v128_t vi3x89AB = wasm_v128_load(i3);
96       i3 += 4;
97 
98       v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
99       v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01));
100       v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, vk11);
101       v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, vk11);
102       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));
103       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));
104 
105       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
106       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
107       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
108       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
109 
110       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk00));
111       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk00));
112       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
113       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10));
114       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk20));
115       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, vk20));
116 
117       vi0x0123 = vi0x4567;
118       vi1x0123 = vi1x4567;
119       vi2x0123 = vi2x4567;
120       vi3x0123 = vi3x4567;
121 
122       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
123       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
124       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
125       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
126 
127       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
128       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk02));
129       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x5678, vk12));
130       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x5678, vk12));
131       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
132       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk22));
133 
134       vi0x4567 = vi0x89AB;
135       vi1x4567 = vi1x89AB;
136       vi2x4567 = vi2x89AB;
137       vi3x4567 = vi3x89AB;
138 
139       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
140       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
141 
142       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
143       v128_t vo1 = wasm_f32x4_pmax(vmin, vo1p0);
144       vo0 = wasm_f32x4_pmin(vmax, vo0);
145       vo1 = wasm_f32x4_pmin(vmax, vo1);
146 
147       wasm_v128_store(o1, vo1);
148       o1 += 4;
149       wasm_v128_store(o0, vo0);
150       o0 += 4;
151     }
152     // Always process the last block of 1..4 pixels.
153     assert(w >= 1 * sizeof(float));
154     assert(w <= 4 * sizeof(float));
155     {
156       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
157       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
158       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
159       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
160 
161       v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
162       v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01));
163       v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, vk11);
164       v128_t vo1p1 = wasm_f32x4_mul(vi2x4567, vk11);
165       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));
166       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));
167 
168       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
169       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
170       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
171       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
172 
173       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x3456, vk00));
174       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x3456, vk00));
175       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
176       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10));
177       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk20));
178       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x3456, vk20));
179 
180       const v128_t vzero = wasm_f32x4_const_splat(0.0f);
181       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
182       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
183       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
184       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
185 
186       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
187       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk02));
188       vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x5678, vk12));
189       vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x5678, vk12));
190       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
191       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk22));
192 
193       vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
194       vo1p0 = wasm_f32x4_add(vo1p0, vo1p1);
195 
196       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
197       v128_t vo1 = wasm_f32x4_pmax(vmin, vo1p0);
198       vo0 = wasm_f32x4_pmin(vmax, vo0);
199       vo1 = wasm_f32x4_pmin(vmax, vo1);
200 
201       if XNN_LIKELY(w == 4 * sizeof(float)) {
202         wasm_v128_store(o1, vo1);
203         o1 += 4;
204         wasm_v128_store(o0, vo0);
205         o0 += 4;
206       } else {
207         if (w & (2 * sizeof(float))) {
208           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0);
209           o1 += 2;
210           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0);
211           o0 += 2;
212 
213           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
214           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
215         }
216         if (w & (1 * sizeof(float))) {
217           *o1 = wasm_f32x4_extract_lane(vo1, 0);
218           o1 += 1;
219           *o0 = wasm_f32x4_extract_lane(vo0, 0);
220           o0 += 1;
221         }
222       }
223     }
224 
225     i0 = (const float*) ((uintptr_t) i2 - input_decrement);
226     i1 = (const float*) ((uintptr_t) i3 - input_decrement);
227     i2 = (const float*) ((uintptr_t) i1 + input_width);
228     i3 = (const float*) ((uintptr_t) i2 + input_width);
229 
230     o0 = o1;
231     o1 = (float*) ((uintptr_t) o0 + input_width);
232 
233     output_height = doz(output_height, 2);
234   } while (output_height != 0);
235 }
236