• 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$ARCH_SUFFIX = "_x86" if X86 else "_arm"
17
18void xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd${ARCH_SUFFIX}_splat_${ROW_TILE}x4${"_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 >= 1);
32  assert(padding_top <= 2);
33
34  const v128_t vmask_even = wasm_v128_load(params->scalar.mask_even);
35  const v128_t vmask_odd  = wasm_v128_load(params->scalar.mask_odd);
36  const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
37  const v128_t vmin = wasm_v32x4_load_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 vw89AB = wasm_v128_load(weights + 8);
42  const v128_t vwCDEF = wasm_v128_load(weights + 12);
43  const v128_t vwGHIJ = wasm_v128_load(weights + 16);
44  const v128_t vwKLMN = wasm_v128_load(weights + 20);
45  const v128_t vwOP = wasm_v64x2_load_splat(weights + 24);
46
47  const v128_t vzero = wasm_f32x4_splat(0.0f);
48
49  const uint32_t padding_top_less_1 = padding_top - 1;
50  const size_t input_decrement = round_up_po2(input_width, 8 * sizeof(float));
51
52  const float* i0 = zero;
53  const float* i1 = (const float*) ((uintptr_t) input - ((-padding_top_less_1) & input_width));
54  const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
55  if XNN_UNPREDICTABLE(padding_top_less_1 != 0) {
56    i1 = zero;
57  }
58  $for M in range(3, 3 + 2 * ROW_TILE):
59    const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);
60
61  $if ROW_TILE > 1:
62    const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
63
64  float* o0 = output;
65  $for M in range(1, ROW_TILE):
66    float* o${M} = (float*) ((uintptr_t) o${M-1} + output_width);
67
68  size_t padded_input_height = input_height + (padding_top_less_1 + 1) + 2 /* padding bottom */;
69  size_t output_height = (padded_input_height - 5 /* kernel size */ + 2 /* subsampling */) / 2;
70  do {
71    $for M in range(3, 3 + 2 * ROW_TILE):
72      if XNN_UNPREDICTABLE(padded_input_height < ${3 + M}) {
73        i${M} = zero;
74        $if M % 2 == 0 and M <= 2 * ROW_TILE + 1:
75          o${M / 2 - 1} = o${M / 2 - 2};
76      }
77
78    $for M in range(3 + 2 * ROW_TILE):
79      v128_t vi${M}x0246 = vzero;
80
81    $for M in range(3 + 2 * ROW_TILE):
82      v128_t vi${M}x1357 = vzero;
83
84    $for M in range(3 + 2 * ROW_TILE):
85      const v128_t vi${M}x89AB = wasm_v128_load(i${M});
86      const v128_t vi${M}xCDEF = wasm_v128_load(i${M} + 4);
87      i${M} += 8;
88
89    $for M in range(3 + 2 * ROW_TILE):
90      v128_t vi${M}x8ACE = wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 0, 2, 4, 6);
91      v128_t vi${M}x9BDF = wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 1, 3, 5, 7);
92
93    size_t w = input_width;
94    for (; w > 8 * sizeof(float); w -= 8 * sizeof(float)) {
95      $for M in range(ROW_TILE):
96        v128_t vo${M}p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
97
98      $for M in range(ROW_TILE):
99        $if ACCUMULATORS > 1:
100          v128_t vo${M}p1 = wasm_f32x4_mul(vi${2*M}x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
101        $else:
102          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M}x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
103
104      $for M in range(ROW_TILE):
105        $if ACCUMULATORS > 2:
106          v128_t vo${M}p2 = wasm_f32x4_mul(vi${2*M+1}x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0));
107        $else:
108          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M+1}x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
109
110      $for M in range(ROW_TILE):
111        $if ACCUMULATORS > 3:
112          v128_t vo${M}p3 = wasm_f32x4_mul(vi${2*M+2}x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1));
113        $else:
114          vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
115
116      $for M in range(ROW_TILE):
117        $if ACCUMULATORS > 4:
118          v128_t vo${M}p4 = wasm_f32x4_mul(vi${2*M+3}x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2));
119        $else:
120          vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
121
122      $for M in range(ROW_TILE):
123        $if ACCUMULATORS > 5:
124          vo${M}p5 = wasm_f32x4_mul(vi${2*M+4}x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3));
125        $else:
126          vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
127
128      $for M in range(ROW_TILE):
129        vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
130
131      $for M in range(ROW_TILE):
132        vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
133
134      $for M in range(ROW_TILE):
135        vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
136
137      $for M in range(ROW_TILE):
138        vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
139
140      $for M in range(ROW_TILE):
141        vo${M}p${11 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${11 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
142
143      $for M in range(3 + 2 * ROW_TILE):
144        const v128_t vi${M}x68AC = wasm_v32x4_shuffle(vi${M}x0246, vi${M}x8ACE, 3, 4, 5, 6);
145        vi${M}x0246 = vi${M}x8ACE;
146
147      $for M in range(ROW_TILE):
148        vo${M}p${12 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${12 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
149
150      $for M in range(ROW_TILE):
151        vo${M}p${13 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${13 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
152
153      $for M in range(ROW_TILE):
154        vo${M}p${14 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${14 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
155
156      $for M in range(ROW_TILE):
157        vo${M}p${15 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${15 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
158
159      $for M in range(ROW_TILE):
160        vo${M}p${16 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${16 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
161
162      $for M in range(3 + 2 * ROW_TILE):
163        const v128_t vi${M}x79BD = wasm_v32x4_shuffle(vi${M}x1357, vi${M}x9BDF, 3, 4, 5, 6);
164        vi${M}x1357 = vi${M}x9BDF;
165
166      $for M in range(3 + 2 * ROW_TILE):
167        const v128_t vi${M}xGHIJ = wasm_v128_load(i${M});
168        const v128_t vi${M}xKLMN = wasm_v128_load(i${M} + 4);
169        i${M} += 8;
170
171      $for M in range(3 + 2 * ROW_TILE):
172        const v128_t vi${M}xGIKM = wasm_v32x4_shuffle(vi${M}xGHIJ, vi${M}xKLMN, 0, 2, 4, 6);
173        const v128_t vi${M}xHJLN = wasm_v32x4_shuffle(vi${M}xGHIJ, vi${M}xKLMN, 1, 3, 5, 7);
174
175      $for M in range(ROW_TILE):
176        vo${M}p${17 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${17 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
177
178      $for M in range(ROW_TILE):
179        vo${M}p${18 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${18 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
180
181      $for M in range(ROW_TILE):
182        vo${M}p${19 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${19 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
183
184      $for M in range(ROW_TILE):
185        vo${M}p${20 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${20 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
186
187      $for M in range(ROW_TILE):
188        vo${M}p${21 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${21 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
189
190      $for M in range(3 + 2 * ROW_TILE):
191        const v128_t vi${M}xACEG = wasm_v32x4_shuffle(vi${M}x8ACE, vi${M}xGIKM, 1, 2, 3, 4);
192        vi${M}x8ACE = vi${M}xGIKM;
193        vi${M}x9BDF = vi${M}xHJLN;
194
195      $for M in range(ROW_TILE):
196        vo${M}p${22 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${22 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
197
198      $for M in range(ROW_TILE):
199        vo${M}p${23 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${23 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
200
201      $for M in range(ROW_TILE):
202        vo${M}p${24 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${24 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
203
204      $for M in range(ROW_TILE):
205        vo${M}p${25 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${25 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
206
207      $for M in range(ROW_TILE):
208        vo${M}p${26 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${26 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
209
210      $if ACCUMULATORS > 1:
211        $ACC_SLICE = 1
212        $while ACC_SLICE < ACCUMULATORS:
213          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
214            $if A + ACC_SLICE < ACCUMULATORS:
215              $for M in range(ROW_TILE):
216                vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
217          $ACC_SLICE *= 2
218
219      $if X86:
220        $for M in range(ROW_TILE):
221          v128_t vo${M} = wasm_v128_bitselect(vmin, vo${M}p0, wasm_f32x4_lt(vo${M}p0, vmin));
222        $for M in range(ROW_TILE):
223          vo${M} = wasm_v128_bitselect(vo${M}, vmax, wasm_f32x4_le(vo${M}, vmax));
224      $else:
225        $for M in range(ROW_TILE):
226          v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
227        $for M in range(ROW_TILE):
228          vo${M} = wasm_f32x4_min(vo${M}, vmax);
229
230      $for M in reversed(range(ROW_TILE)):
231        wasm_v128_store(o${M}, vo${M}); o${M} += 4;
232    }
233    // Last block has 1-8 pixels to process.
234    assert(w <= 8 * sizeof(float));
235    assert(w >= 1 * sizeof(float));
236    {
237      $for M in range(ROW_TILE):
238        v128_t vo${M}p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
239
240      $for M in range(3 + 2 * ROW_TILE):
241        vi${M}x8ACE = wasm_v128_and(vmask_even, vi${M}x8ACE);
242
243      $for M in range(3 + 2 * ROW_TILE):
244        vi${M}x9BDF = wasm_v128_and(vmask_odd, vi${M}x9BDF);
245
246      $for M in range(ROW_TILE):
247        $if ACCUMULATORS > 1:
248          v128_t vo${M}p1 = wasm_f32x4_mul(vi${2*M}x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3));
249        $else:
250          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M}x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
251
252      $for M in range(ROW_TILE):
253        $if ACCUMULATORS > 2:
254          v128_t vo${M}p2 = wasm_f32x4_mul(vi${2*M+1}x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0));
255        $else:
256          vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M+1}x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)));
257
258      $for M in range(ROW_TILE):
259        $if ACCUMULATORS > 3:
260          v128_t vo${M}p3 = wasm_f32x4_mul(vi${2*M+2}x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1));
261        $else:
262          vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x8ACE, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 1, 1, 1, 1)));
263
264      $for M in range(ROW_TILE):
265        $if ACCUMULATORS > 4:
266          v128_t vo${M}p4 = wasm_f32x4_mul(vi${2*M+3}x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2));
267        $else:
268          vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}x8ACE, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 2, 2, 2, 2)));
269
270      $for M in range(ROW_TILE):
271        $if ACCUMULATORS > 5:
272          vo${M}p5 = wasm_f32x4_mul(vi${2*M+4}x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3));
273        $else:
274          vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}x8ACE, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 3, 3, 3, 3)));
275
276      $for M in range(ROW_TILE):
277        vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
278
279      $for M in range(ROW_TILE):
280        vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1, 1)));
281
282      $for M in range(ROW_TILE):
283        vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x9BDF, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2)));
284
285      $for M in range(ROW_TILE):
286        vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}x9BDF, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 3, 3, 3, 3)));
287
288      $for M in range(ROW_TILE):
289        vo${M}p${11 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${11 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0)));
290
291      $for M in range(3 + 2 * ROW_TILE):
292        const v128_t vi${M}x68AC = wasm_v32x4_shuffle(vi${M}x0246, vi${M}x8ACE, 3, 4, 5, 6);
293
294      $for M in range(ROW_TILE):
295        vo${M}p${12 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${12 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x68AC, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
296
297      $for M in range(ROW_TILE):
298        vo${M}p${13 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${13 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x68AC, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
299
300      $for M in range(ROW_TILE):
301        vo${M}p${14 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${14 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3, 3)));
302
303      $for M in range(ROW_TILE):
304        vo${M}p${15 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${15 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}x68AC, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 0, 0, 0, 0)));
305
306      $for M in range(ROW_TILE):
307        vo${M}p${16 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${16 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}x68AC, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1, 1)));
308
309      $for M in range(3 + 2 * ROW_TILE):
310        const v128_t vi${M}x79BD = wasm_v32x4_shuffle(vi${M}x1357, vi${M}x9BDF, 3, 4, 5, 6);
311
312      $for M in range(ROW_TILE):
313        vo${M}p${17 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${17 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
314
315      $for M in range(ROW_TILE):
316        vo${M}p${18 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${18 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x79BD, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
317
318      $for M in range(ROW_TILE):
319        vo${M}p${19 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${19 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x79BD, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 0, 0, 0, 0)));
320
321      $for M in range(ROW_TILE):
322        vo${M}p${20 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${20 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}x79BD, wasm_v32x4_shuffle(vwGHIJ, vwGHIJ, 1, 1, 1, 1)));
323
324      $for M in range(ROW_TILE):
325        vo${M}p${21 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${21 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}x79BD, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 2, 2, 2, 2)));
326
327      $for M in range(3 + 2 * ROW_TILE):
328        const v128_t vi${M}xACEG = wasm_v32x4_shuffle(vi${M}x8ACE, vzero, 1, 2, 3, 4);
329
330      $for M in range(ROW_TILE):
331        vo${M}p${22 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${22 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}xACEG, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
332
333      $for M in range(ROW_TILE):
334        vo${M}p${23 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${23 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2, 2)));
335
336      $for M in range(ROW_TILE):
337        vo${M}p${24 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${24 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}xACEG, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3, 3)));
338
339      $for M in range(ROW_TILE):
340        vo${M}p${25 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${25 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+3}xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0, 0)));
341
342      $for M in range(ROW_TILE):
343        vo${M}p${26 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${26 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+4}xACEG, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1)));
344
345      $if ACCUMULATORS > 1:
346        $ACC_SLICE = 1
347        $while ACC_SLICE < ACCUMULATORS:
348          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
349            $if A + ACC_SLICE < ACCUMULATORS:
350              $for M in range(ROW_TILE):
351                vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
352          $ACC_SLICE *= 2
353
354      $if X86:
355        $for M in range(ROW_TILE):
356          v128_t vo${M} = wasm_v128_bitselect(vmin, vo${M}p0, wasm_f32x4_lt(vo${M}p0, vmin));
357        $for M in range(ROW_TILE):
358          vo${M} = wasm_v128_bitselect(vo${M}, vmax, wasm_f32x4_le(vo${M}, vmax));
359      $else:
360        $for M in range(ROW_TILE):
361          v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
362        $for M in range(ROW_TILE):
363          vo${M} = wasm_f32x4_min(vo${M}, vmax);
364
365      size_t w_tmp = (w + 1 * sizeof(float)) / (2 * sizeof(float));
366      if XNN_LIKELY(w_tmp >= 4) {
367        $for M in reversed(range(ROW_TILE)):
368          wasm_v128_store(o${M}, vo${M}); o${M} += 4;
369      } else {
370        if (w_tmp & 2) {
371          $for M in reversed(range(ROW_TILE)):
372            *((double*) o${M}) = wasm_f64x2_extract_lane(vo${M}, 0); o${M} += 2;
373
374          $for M in range(ROW_TILE):
375            vo${M} = wasm_v32x4_shuffle(vo${M}, vo${M}, 2, 3, 0, 1);
376        }
377        if (w_tmp & 1) {
378          $for M in reversed(range(ROW_TILE)):
379            *o${M} = wasm_f32x4_extract_lane(vo${M}, 0); o${M} += 1;
380        }
381      }
382    }
383
384    i0 = (const float*) ((uintptr_t) i${2 * ROW_TILE} - input_decrement);
385    i1 = (const float*) ((uintptr_t) i${2 * ROW_TILE + 1} - input_decrement);
386    i2 = (const float*) ((uintptr_t) i${2 * ROW_TILE + 2} - input_decrement);
387    $for M in range(3, 3 + 2 * ROW_TILE):
388      i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);
389
390    $if ROW_TILE > 1:
391      o0 = o${ROW_TILE - 1};
392      $for M in range(1, ROW_TILE):
393        o${M} = (float*) ((uintptr_t) o${M-1} + output_width);
394
395    $if ROW_TILE > 1:
396      output_height = doz(output_height, ${ROW_TILE});
397      padded_input_height = doz(padded_input_height, ${ROW_TILE * 2});
398    $else:
399      output_height -= 1;
400      padded_input_height -= 2;
401  } while (output_height != 0);
402}
403