• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1// Copyright 2020 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6$assert ROW_TILE >= 1
7$assert ACCUMULATORS >= 1
8#include <assert.h>
9
10#include <wasm_simd128.h>
11
12#include <xnnpack/dwconv.h>
13#include <xnnpack/math.h>
14
15
16$ARCH_SUFFIX = "_x86" if X86 else "_arm"
17
18void xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd${ARCH_SUFFIX}_loadsplat_${ROW_TILE}x4${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}(
19    size_t input_height,
20    size_t input_width,
21    const float* input,
22    const float* weights,
23    const float* zero,
24    float* output,
25    uint32_t padding_top,
26    const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
27{
28  assert(input_height != 0);
29  assert(input_width != 0);
30  assert(input_width % sizeof(float) == 0);
31  assert(padding_top == 2);
32
33  const v128_t vmask = wasm_v128_load(params->scalar.mask);
34  const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
35  const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
36
37  const v128_t vw0123 = wasm_v128_load(weights);
38  const v128_t vw4567 = wasm_v128_load(weights + 4);
39  const v128_t vw89AB = wasm_v128_load(weights + 8);
40  const v128_t vwCDEF = wasm_v128_load(weights + 12);
41  const v128_t vwGHIJ = wasm_v128_load(weights + 16);
42  const v128_t vwKLMN = wasm_v128_load(weights + 20);
43  const v128_t vwOP = wasm_v64x2_load_splat(weights + 24);
44  const v128_t vbias = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
45  const v128_t vk00 =  wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1);
46  const v128_t vk01 =  wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2);
47  const v128_t vk02 =  wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3);
48  const v128_t vk03 =  wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0);
49  const v128_t vk04 =  wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1);
50  const v128_t vk10 =  wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2);
51  const v128_t vk11 =  wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3);
52  const v128_t vk12 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0);
53  const v128_t vk13 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1);
54  const v128_t vk14 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2);
55  const v128_t vk20 =  wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3);
56  const v128_t vk21 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0);
57  const v128_t vk22 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1);
58  const v128_t vk23 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2);
59  const v128_t vk24 =  wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3);
60  const v128_t vk30 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0);
61  const v128_t vk31 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1);
62  const v128_t vk32 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2);
63  const v128_t vk33 =  wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3);
64  const v128_t vk34 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0);
65  const v128_t vk40 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1);
66  const v128_t vk41 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2);
67  const v128_t vk42 =  wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3);
68  const v128_t vk43 =  wasm_v32x4_shuffle(vwOP,   vwOP,   0, 0, 0, 0);
69  const v128_t vk44 =  wasm_v32x4_shuffle(vwOP,   vwOP,   1, 1, 1, 1);
70
71  const v128_t vzero = wasm_f32x4_splat(0.0f);
72
73  const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
74
75  const float* i0 = zero;
76  const float* i1 = zero;
77  const float* i2 = input;
78  $for M in range(3, 4 + ROW_TILE):
79    const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);
80
81  float* o0 = output;
82  $for M in range(1, ROW_TILE):
83    float* o${M} = (float*) ((uintptr_t) o${M-1} + input_width);
84
85  size_t output_height = input_height;
86  do {
87    $for M in range(2, 3 + ROW_TILE):
88      if XNN_UNPREDICTABLE(output_height < ${M}) {
89        i${M+1} = zero;
90        $if M <= ROW_TILE:
91          o${M-1} = o${M-2};
92      }
93
94    $for M in range(4 + ROW_TILE):
95      v128_t vi${M}x0123 = vzero;
96
97    $for M in range(4 + ROW_TILE):
98      v128_t vi${M}x4567 = wasm_v128_load(i${M}); i${M} += 4;
99
100    size_t w = input_width;
101    for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
102      $for M in range(ROW_TILE):
103        v128_t vo${M}p0 = vbias;
104
105      $for M in range(4 + ROW_TILE):
106        const v128_t vi${M}x89AB = wasm_v128_load(i${M}); i${M} += 4;
107
108      $for M in range(ROW_TILE):
109        $if ACCUMULATORS > 1:
110          v128_t vo${M}p1 = wasm_f32x4_mul(vi${M}x4567, vk02);
111        $else:
112          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${M}x4567, vk02));
113
114      $for M in range(ROW_TILE):
115        $if ACCUMULATORS > 2:
116          v128_t vo${M}p2 = wasm_f32x4_mul(vi${M+1}x4567, vk12);
117        $else:
118          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${M+1}x4567, vk12));
119
120      $for M in range(ROW_TILE):
121        $if ACCUMULATORS > 3:
122          v128_t vo${M}p3 = wasm_f32x4_mul(vi${M+2}x4567, vk22);
123        $else:
124          vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x4567, vk22));
125
126      $for M in range(ROW_TILE):
127        $if ACCUMULATORS > 4:
128          v128_t vo${M}p4 = wasm_f32x4_mul(vi${M+3}x4567, vk32);
129        $else:
130          vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x4567, vk32));
131
132      $for M in range(ROW_TILE):
133        $if ACCUMULATORS > 6:
134          v128_t vo${M}p5 = wasm_f32x4_mul(vi${M+4}x4567, vk42);
135        $else:
136          vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x4567, vk42));
137
138      $for M in range(4 + ROW_TILE):
139        const v128_t vi${M}x3456 = wasm_v32x4_shuffle(vi${M}x0123, vi${M}x4567, 3, 4, 5, 6);
140
141      $for M in range(ROW_TILE):
142        vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x3456, vk01));
143
144      $for M in range(ROW_TILE):
145        vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x3456, vk11));
146
147      $for M in range(ROW_TILE):
148        vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x3456, vk21));
149
150      $for M in range(ROW_TILE):
151        vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x3456, vk31));
152
153      $for M in range(ROW_TILE):
154        vo${M}p${11 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${11 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x3456, vk41));
155
156      $for M in range(4 + ROW_TILE):
157        const v128_t vi${M}x2345 = wasm_v32x4_shuffle(vi${M}x0123, vi${M}x4567, 2, 3, 4, 5);
158        vi${M}x0123 = vi${M}x4567;
159
160      $for M in range(ROW_TILE):
161        vo${M}p${12 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${12 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x2345, vk00));
162
163      $for M in range(ROW_TILE):
164        vo${M}p${13 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${13 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x2345, vk10));
165
166      $for M in range(ROW_TILE):
167        vo${M}p${14 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${14 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x2345, vk20));
168
169      $for M in range(ROW_TILE):
170        vo${M}p${15 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${15 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x2345, vk30));
171
172      $for M in range(ROW_TILE):
173        vo${M}p${16 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${16 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x2345, vk40));
174
175      $for M in range(4 + ROW_TILE):
176        const v128_t vi${M}x5678 = wasm_v32x4_shuffle(vi${M}x4567, vi${M}x89AB, 1, 2, 3, 4);
177
178      $for M in range(ROW_TILE):
179        vo${M}p${17 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${17 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x5678, vk03));
180
181      $for M in range(ROW_TILE):
182        vo${M}p${18 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${18 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x5678, vk13));
183
184      $for M in range(ROW_TILE):
185        vo${M}p${19 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${19 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x5678, vk23));
186
187      $for M in range(ROW_TILE):
188        vo${M}p${20 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${20 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x5678, vk33));
189
190      $for M in range(ROW_TILE):
191        vo${M}p${21 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${21 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x5678, vk43));
192
193      $for M in range(4 + ROW_TILE):
194        const v128_t vi${M}x6789 = wasm_v32x4_shuffle(vi${M}x4567, vi${M}x89AB, 2, 3, 4, 5);
195        vi${M}x4567 = vi${M}x89AB;
196
197      $for M in range(ROW_TILE):
198        vo${M}p${22 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${22 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x6789, vk04));
199
200      $for M in range(ROW_TILE):
201        vo${M}p${23 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${23 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x6789, vk14));
202
203      $for M in range(ROW_TILE):
204        vo${M}p${24 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${24 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x6789, vk24));
205
206      $for M in range(ROW_TILE):
207        vo${M}p${25 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${25 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x6789, vk34));
208
209      $for M in range(ROW_TILE):
210        vo${M}p${26 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${26 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x6789, vk44));
211
212      $if ACCUMULATORS > 1:
213        $ACC_SLICE = 1
214        $while ACC_SLICE < ACCUMULATORS:
215          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
216            $if A + ACC_SLICE < ACCUMULATORS:
217              $for M in range(ROW_TILE):
218                vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
219          $ACC_SLICE *= 2
220
221      $if X86:
222        $for M in range(ROW_TILE):
223          v128_t vo${M} = wasm_v128_bitselect(vmin, vo${M}p0, wasm_f32x4_lt(vo${M}p0, vmin));
224        $for M in range(ROW_TILE):
225          vo${M} = wasm_v128_bitselect(vo${M}, vmax, wasm_f32x4_le(vo${M}, vmax));
226      $else:
227        $for M in range(ROW_TILE):
228          v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
229        $for M in range(ROW_TILE):
230          vo${M} = wasm_f32x4_min(vo${M}, vmax);
231
232      $for M in reversed(range(ROW_TILE)):
233        wasm_v128_store(o${M}, vo${M}); o${M} += 4;
234    }
235    // Always process the last block of 5..8 pixels.
236    if XNN_LIKELY(w > 4 * sizeof(float)) {
237      $for M in range(ROW_TILE):
238        v128_t vo${M}p0 = vbias;
239
240      $for M in range(4 + ROW_TILE):
241        v128_t vi${M}x89AB = wasm_v128_load(i${M}); i${M} += 4;
242
243      $for M in range(4 + ROW_TILE):
244        vi${M}x89AB = wasm_v128_and(vmask, vi${M}x89AB);
245
246      $for M in range(ROW_TILE):
247        $if ACCUMULATORS > 1:
248          v128_t vo${M}p1 = wasm_f32x4_mul(vi${M}x4567, vk02);
249        $else:
250          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${M}x4567, vk02));
251
252      $for M in range(ROW_TILE):
253        $if ACCUMULATORS > 2:
254          v128_t vo${M}p2 = wasm_f32x4_mul(vi${M+1}x4567, vk12);
255        $else:
256          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${M+1}x4567, vk12));
257
258      $for M in range(ROW_TILE):
259        $if ACCUMULATORS > 3:
260          v128_t vo${M}p3 = wasm_f32x4_mul(vi${M+2}x4567, vk22);
261        $else:
262          vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x4567, vk22));
263
264      $for M in range(ROW_TILE):
265        $if ACCUMULATORS > 4:
266          v128_t vo${M}p4 = wasm_f32x4_mul(vi${M+3}x4567, vk32);
267        $else:
268          vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x4567, vk32));
269
270      $for M in range(ROW_TILE):
271        $if ACCUMULATORS > 6:
272          v128_t vo${M}p5 = wasm_f32x4_mul(vi${M+4}x4567, vk42);
273        $else:
274          vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x4567, vk42));
275
276      $for M in range(4 + ROW_TILE):
277        const v128_t vi${M}x3456 = wasm_v32x4_shuffle(vi${M}x0123, vi${M}x4567, 3, 4, 5, 6);
278
279      $for M in range(ROW_TILE):
280        vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x3456, vk01));
281
282      $for M in range(ROW_TILE):
283        vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x3456, vk11));
284
285      $for M in range(ROW_TILE):
286        vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x3456, vk21));
287
288      $for M in range(ROW_TILE):
289        vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x3456, vk31));
290
291      $for M in range(ROW_TILE):
292        vo${M}p${11 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${11 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x3456, vk41));
293
294      $for M in range(4 + ROW_TILE):
295        const v128_t vi${M}x2345 = wasm_v32x4_shuffle(vi${M}x0123, vi${M}x4567, 2, 3, 4, 5);
296        vi${M}x0123 = vi${M}x4567;
297
298      $for M in range(ROW_TILE):
299        vo${M}p${12 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${12 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x2345, vk00));
300
301      $for M in range(ROW_TILE):
302        vo${M}p${13 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${13 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x2345, vk10));
303
304      $for M in range(ROW_TILE):
305        vo${M}p${14 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${14 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x2345, vk20));
306
307      $for M in range(ROW_TILE):
308        vo${M}p${15 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${15 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x2345, vk30));
309
310      $for M in range(ROW_TILE):
311        vo${M}p${16 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${16 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x2345, vk40));
312
313      $for M in range(4 + ROW_TILE):
314        const v128_t vi${M}x5678 = wasm_v32x4_shuffle(vi${M}x4567, vi${M}x89AB, 1, 2, 3, 4);
315
316      $for M in range(ROW_TILE):
317        vo${M}p${17 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${17 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x5678, vk03));
318
319      $for M in range(ROW_TILE):
320        vo${M}p${18 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${18 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x5678, vk13));
321
322      $for M in range(ROW_TILE):
323        vo${M}p${19 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${19 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x5678, vk23));
324
325      $for M in range(ROW_TILE):
326        vo${M}p${20 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${20 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x5678, vk33));
327
328      $for M in range(ROW_TILE):
329        vo${M}p${21 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${21 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x5678, vk43));
330
331      $for M in range(4 + ROW_TILE):
332        const v128_t vi${M}x6789 = wasm_v32x4_shuffle(vi${M}x4567, vi${M}x89AB, 2, 3, 4, 5);
333        vi${M}x4567 = vi${M}x89AB;
334
335      $for M in range(ROW_TILE):
336        vo${M}p${22 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${22 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x6789, vk04));
337
338      $for M in range(ROW_TILE):
339        vo${M}p${23 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${23 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x6789, vk14));
340
341      $for M in range(ROW_TILE):
342        vo${M}p${24 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${24 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x6789, vk24));
343
344      $for M in range(ROW_TILE):
345        vo${M}p${25 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${25 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x6789, vk34));
346
347      $for M in range(ROW_TILE):
348        vo${M}p${26 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${26 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x6789, vk44));
349
350      $if ACCUMULATORS > 1:
351        $ACC_SLICE = 1
352        $while ACC_SLICE < ACCUMULATORS:
353          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
354            $if A + ACC_SLICE < ACCUMULATORS:
355              $for M in range(ROW_TILE):
356                vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
357          $ACC_SLICE *= 2
358
359      $if X86:
360        $for M in range(ROW_TILE):
361          v128_t vo${M} = wasm_v128_bitselect(vmin, vo${M}p0, wasm_f32x4_lt(vo${M}p0, vmin));
362        $for M in range(ROW_TILE):
363          vo${M} = wasm_v128_bitselect(vo${M}, vmax, wasm_f32x4_le(vo${M}, vmax));
364      $else:
365        $for M in range(ROW_TILE):
366          v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
367        $for M in range(ROW_TILE):
368          vo${M} = wasm_f32x4_min(vo${M}, vmax);
369
370      $for M in reversed(range(ROW_TILE)):
371        wasm_v128_store(o${M}, vo${M}); o${M} += 4;
372
373      w -= 4 * sizeof(float);
374    }
375    assert(w >= 1 * sizeof(float));
376    assert(w <= 4 * sizeof(float));
377    {
378      $for M in range(ROW_TILE):
379        v128_t vo${M}p0 = vbias;
380
381      $for M in range(4 + ROW_TILE):
382        vi${M}x4567 = wasm_v128_and(vmask, vi${M}x4567);
383
384      $for M in range(ROW_TILE):
385        $if ACCUMULATORS > 1:
386          v128_t vo${M}p1 = wasm_f32x4_mul(vi${M}x4567, vk02);
387        $else:
388          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${M}x4567, vk02));
389
390      $for M in range(ROW_TILE):
391        $if ACCUMULATORS > 2:
392          v128_t vo${M}p2 = wasm_f32x4_mul(vi${M+1}x4567, vk12);
393        $else:
394          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${M+1}x4567, vk12));
395
396      $for M in range(ROW_TILE):
397        $if ACCUMULATORS > 3:
398          v128_t vo${M}p3 = wasm_f32x4_mul(vi${M+2}x4567, vk22);
399        $else:
400          vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x4567, vk22));
401
402      $for M in range(ROW_TILE):
403        $if ACCUMULATORS > 4:
404          v128_t vo${M}p4 = wasm_f32x4_mul(vi${M+3}x4567, vk32);
405        $else:
406          vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x4567, vk32));
407
408      $for M in range(ROW_TILE):
409        $if ACCUMULATORS > 6:
410          v128_t vo${M}p5 = wasm_f32x4_mul(vi${M+4}x4567, vk42);
411        $else:
412          vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x4567, vk42));
413
414      $for M in range(4 + ROW_TILE):
415        const v128_t vi${M}x3456 = wasm_v32x4_shuffle(vi${M}x0123, vi${M}x4567, 3, 4, 5, 6);
416
417      $for M in range(ROW_TILE):
418        vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x3456, vk01));
419
420      $for M in range(ROW_TILE):
421        vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x3456, vk11));
422
423      $for M in range(ROW_TILE):
424        vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x3456, vk21));
425
426      $for M in range(ROW_TILE):
427        vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x3456, vk31));
428
429      $for M in range(ROW_TILE):
430        vo${M}p${11 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${11 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x3456, vk41));
431
432      $for M in range(4 + ROW_TILE):
433        const v128_t vi${M}x2345 = wasm_v32x4_shuffle(vi${M}x0123, vi${M}x4567, 2, 3, 4, 5);
434
435      $for M in range(ROW_TILE):
436        vo${M}p${12 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${12 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x2345, vk00));
437
438      $for M in range(ROW_TILE):
439        vo${M}p${13 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${13 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x2345, vk10));
440
441      $for M in range(ROW_TILE):
442        vo${M}p${14 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${14 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x2345, vk20));
443
444      $for M in range(ROW_TILE):
445        vo${M}p${15 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${15 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x2345, vk30));
446
447      $for M in range(ROW_TILE):
448        vo${M}p${16 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${16 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x2345, vk40));
449
450      $for M in range(4 + ROW_TILE):
451        const v128_t vi${M}x5678 = wasm_v32x4_shuffle(vi${M}x4567, vzero, 1, 2, 3, 4);
452
453      $for M in range(ROW_TILE):
454        vo${M}p${17 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${17 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x5678, vk03));
455
456      $for M in range(ROW_TILE):
457        vo${M}p${18 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${18 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x5678, vk13));
458
459      $for M in range(ROW_TILE):
460        vo${M}p${19 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${19 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x5678, vk23));
461
462      $for M in range(ROW_TILE):
463        vo${M}p${20 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${20 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x5678, vk33));
464
465      $for M in range(ROW_TILE):
466        vo${M}p${21 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${21 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x5678, vk43));
467
468      $for M in range(4 + ROW_TILE):
469        const v128_t vi${M}x6789 = wasm_v32x4_shuffle(vi${M}x5678, vzero, 1, 2, 3, 4);
470
471      $for M in range(ROW_TILE):
472        vo${M}p${22 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${22 % ACCUMULATORS}, wasm_f32x4_mul(vi${M}x6789, vk04));
473
474      $for M in range(ROW_TILE):
475        vo${M}p${23 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${23 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+1}x6789, vk14));
476
477      $for M in range(ROW_TILE):
478        vo${M}p${24 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${24 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+2}x6789, vk24));
479
480      $for M in range(ROW_TILE):
481        vo${M}p${25 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${25 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+3}x6789, vk34));
482
483      $for M in range(ROW_TILE):
484        vo${M}p${26 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${26 % ACCUMULATORS}, wasm_f32x4_mul(vi${M+4}x6789, vk44));
485
486      $if ACCUMULATORS > 1:
487        $ACC_SLICE = 1
488        $while ACC_SLICE < ACCUMULATORS:
489          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
490            $if A + ACC_SLICE < ACCUMULATORS:
491              $for M in range(ROW_TILE):
492                vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
493          $ACC_SLICE *= 2
494
495      $if X86:
496        $for M in range(ROW_TILE):
497          v128_t vo${M} = wasm_v128_bitselect(vmin, vo${M}p0, wasm_f32x4_lt(vo${M}p0, vmin));
498        $for M in range(ROW_TILE):
499          vo${M} = wasm_v128_bitselect(vo${M}, vmax, wasm_f32x4_le(vo${M}, vmax));
500      $else:
501        $for M in range(ROW_TILE):
502          v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
503        $for M in range(ROW_TILE):
504          vo${M} = wasm_f32x4_min(vo${M}, vmax);
505
506      if XNN_LIKELY(w & (4 * sizeof(float))) {
507        $for M in reversed(range(ROW_TILE)):
508          wasm_v128_store(o${M}, vo${M}); o${M} += 4;
509      } else {
510        if (w & (2 * sizeof(float))) {
511          $for M in reversed(range(ROW_TILE)):
512            *((double*) o${M}) = wasm_f64x2_extract_lane(vo${M}, 0); o${M} += 2;
513
514          $for M in range(ROW_TILE):
515            vo${M} = wasm_v32x4_shuffle(vo${M}, vo${M}, 2, 3, 0, 1);
516        }
517        if (w & (1 * sizeof(float))) {
518          $for M in reversed(range(ROW_TILE)):
519            *o${M} = wasm_f32x4_extract_lane(vo${M}, 0); o${M} += 1;
520        }
521      }
522    }
523
524    i0 = (const float*) ((uintptr_t) i${ROW_TILE} - input_decrement);
525    i1 = (const float*) ((uintptr_t) i${ROW_TILE+1} - input_decrement);
526    $for M in range(2, 4 + ROW_TILE):
527      i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);
528
529    $if ROW_TILE > 1:
530      o0 = o${ROW_TILE - 1};
531      $for M in range(1, ROW_TILE):
532        o${M} = (float*) ((uintptr_t) o${M-1} + input_width);
533
534    $if ROW_TILE > 1:
535      output_height = doz(output_height, ${ROW_TILE});
536  } while (${"--" if ROW_TILE == 1 else ""}output_height != 0);
537}
538