xref: /aosp_15_r20/external/XNNPACK/src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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_4x4(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_4x4(
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   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
60   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
61 
62   float* o0 = output;
63   float* o1 = (float*) ((uintptr_t) o0 + input_width);
64   float* o2 = (float*) ((uintptr_t) o1 + input_width);
65   float* o3 = (float*) ((uintptr_t) o2 + input_width);
66 
67   size_t output_height = input_height;
68   do {
69     if XNN_UNPREDICTABLE(output_height < 2) {
70       i2 = zero;
71       o1 = o0;
72     }
73     if XNN_UNPREDICTABLE(output_height < 3) {
74       i3 = zero;
75       o2 = o1;
76     }
77     if XNN_UNPREDICTABLE(output_height < 4) {
78       i4 = zero;
79       o3 = o2;
80     }
81     if XNN_UNPREDICTABLE(output_height < 5) {
82       i5 = zero;
83     }
84 
85     v128_t vi0x0123 = wasm_f32x4_const_splat(0.0f);
86     v128_t vi1x0123 = wasm_f32x4_const_splat(0.0f);
87     v128_t vi2x0123 = wasm_f32x4_const_splat(0.0f);
88     v128_t vi3x0123 = wasm_f32x4_const_splat(0.0f);
89     v128_t vi4x0123 = wasm_f32x4_const_splat(0.0f);
90     v128_t vi5x0123 = wasm_f32x4_const_splat(0.0f);
91 
92     v128_t vi0x4567 = wasm_v128_load(i0);
93     i0 += 4;
94     v128_t vi1x4567 = wasm_v128_load(i1);
95     i1 += 4;
96     v128_t vi2x4567 = wasm_v128_load(i2);
97     i2 += 4;
98     v128_t vi3x4567 = wasm_v128_load(i3);
99     i3 += 4;
100     v128_t vi4x4567 = wasm_v128_load(i4);
101     i4 += 4;
102     v128_t vi5x4567 = wasm_v128_load(i5);
103     i5 += 4;
104 
105     size_t w = input_width;
106     for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
107       const v128_t vi0x89AB = wasm_v128_load(i0);
108       i0 += 4;
109       const v128_t vi1x89AB = wasm_v128_load(i1);
110       i1 += 4;
111       const v128_t vi2x89AB = wasm_v128_load(i2);
112       i2 += 4;
113       const v128_t vi3x89AB = wasm_v128_load(i3);
114       i3 += 4;
115       const v128_t vi4x89AB = wasm_v128_load(i4);
116       i4 += 4;
117       const v128_t vi5x89AB = wasm_v128_load(i5);
118       i5 += 4;
119 
120       v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
121       v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01));
122       v128_t vo2p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi2x4567, vk01));
123       v128_t vo3p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi3x4567, vk01));
124       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11));
125       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11));
126       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11));
127       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk11));
128       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));
129       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));
130       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk21));
131       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk21));
132 
133       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
134       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
135       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
136       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
137       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
138       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
139 
140       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00));
141       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00));
142       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00));
143       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, vk00));
144       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
145       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10));
146       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk10));
147       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, vk10));
148       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20));
149       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk20));
150       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk20));
151       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20));
152 
153       vi0x0123 = vi0x4567;
154       vi1x0123 = vi1x4567;
155       vi2x0123 = vi2x4567;
156       vi3x0123 = vi3x4567;
157       vi4x0123 = vi4x4567;
158       vi5x0123 = vi5x4567;
159 
160       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
161       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
162       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
163       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
164       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
165       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
166 
167       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
168       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk02));
169       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk02));
170       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, vk02));
171       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk12));
172       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk12));
173       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk12));
174       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk12));
175       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
176       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk22));
177       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk22));
178       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk22));
179 
180       vi0x4567 = vi0x89AB;
181       vi1x4567 = vi1x89AB;
182       vi2x4567 = vi2x89AB;
183       vi3x4567 = vi3x89AB;
184       vi4x4567 = vi4x89AB;
185       vi5x4567 = vi5x89AB;
186 
187 
188       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
189       v128_t vo1 = wasm_f32x4_pmax(vmin, vo1p0);
190       v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0);
191       v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0);
192       vo0 = wasm_f32x4_pmin(vmax, vo0);
193       vo1 = wasm_f32x4_pmin(vmax, vo1);
194       vo2 = wasm_f32x4_pmin(vmax, vo2);
195       vo3 = wasm_f32x4_pmin(vmax, vo3);
196 
197       wasm_v128_store(o3, vo3);
198       o3 += 4;
199       wasm_v128_store(o2, vo2);
200       o2 += 4;
201       wasm_v128_store(o1, vo1);
202       o1 += 4;
203       wasm_v128_store(o0, vo0);
204       o0 += 4;
205     }
206     // Always process the last block of 1..4 pixels.
207     assert(w >= 1 * sizeof(float));
208     assert(w <= 4 * sizeof(float));
209     {
210       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
211       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
212       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
213       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
214       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
215       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
216 
217       v128_t vo0p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi0x4567, vk01));
218       v128_t vo1p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi1x4567, vk01));
219       v128_t vo2p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi2x4567, vk01));
220       v128_t vo3p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi3x4567, vk01));
221       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11));
222       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11));
223       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11));
224       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk11));
225       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, vk21));
226       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, vk21));
227       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, vk21));
228       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, vk21));
229 
230       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
231       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
232       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
233       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
234       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
235       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
236 
237       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00));
238       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00));
239       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00));
240       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, vk00));
241       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, vk10));
242       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10));
243       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, vk10));
244       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, vk10));
245       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20));
246       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, vk20));
247       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, vk20));
248       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20));
249 
250       const v128_t vzero = wasm_f32x4_const_splat(0.0f);
251       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
252       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
253       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
254       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
255       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
256       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
257 
258       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02));
259       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk02));
260       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk02));
261       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, vk02));
262       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, vk12));
263       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, vk12));
264       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, vk12));
265       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, vk12));
266       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk22));
267       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk22));
268       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk22));
269       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk22));
270 
271 
272       v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
273       v128_t vo1 = wasm_f32x4_pmax(vmin, vo1p0);
274       v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0);
275       v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0);
276       vo0 = wasm_f32x4_pmin(vmax, vo0);
277       vo1 = wasm_f32x4_pmin(vmax, vo1);
278       vo2 = wasm_f32x4_pmin(vmax, vo2);
279       vo3 = wasm_f32x4_pmin(vmax, vo3);
280 
281       if XNN_LIKELY(w == 4 * sizeof(float)) {
282         wasm_v128_store(o3, vo3);
283         o3 += 4;
284         wasm_v128_store(o2, vo2);
285         o2 += 4;
286         wasm_v128_store(o1, vo1);
287         o1 += 4;
288         wasm_v128_store(o0, vo0);
289         o0 += 4;
290       } else {
291         if (w & (2 * sizeof(float))) {
292           *((double*) o3) = wasm_f64x2_extract_lane(vo3, 0);
293           o3 += 2;
294           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0);
295           o2 += 2;
296           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0);
297           o1 += 2;
298           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0);
299           o0 += 2;
300 
301           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
302           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
303           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
304           vo3 = wasm_v32x4_shuffle(vo3, vo3, 2, 3, 0, 1);
305         }
306         if (w & (1 * sizeof(float))) {
307           *o3 = wasm_f32x4_extract_lane(vo3, 0);
308           o3 += 1;
309           *o2 = wasm_f32x4_extract_lane(vo2, 0);
310           o2 += 1;
311           *o1 = wasm_f32x4_extract_lane(vo1, 0);
312           o1 += 1;
313           *o0 = wasm_f32x4_extract_lane(vo0, 0);
314           o0 += 1;
315         }
316       }
317     }
318 
319     i0 = (const float*) ((uintptr_t) i4 - input_decrement);
320     i1 = (const float*) ((uintptr_t) i5 - input_decrement);
321     i2 = (const float*) ((uintptr_t) i1 + input_width);
322     i3 = (const float*) ((uintptr_t) i2 + input_width);
323     i4 = (const float*) ((uintptr_t) i3 + input_width);
324     i5 = (const float*) ((uintptr_t) i4 + input_width);
325 
326     o0 = o3;
327     o1 = (float*) ((uintptr_t) o0 + input_width);
328     o2 = (float*) ((uintptr_t) o1 + input_width);
329     o3 = (float*) ((uintptr_t) o2 + input_width);
330 
331     output_height = doz(output_height, 4);
332   } while (output_height != 0);
333 }
334