xref: /aosp_15_r20/external/XNNPACK/src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-dwconv2d-chw/3x3p1-wasmsimd-splat.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <wasm_simd128.h>
13*4bdc9457SAndroid Build Coastguard Worker 
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/dwconv.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker 
18*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4(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*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4(
20*4bdc9457SAndroid Build Coastguard Worker     size_t input_height,
21*4bdc9457SAndroid Build Coastguard Worker     size_t input_width,
22*4bdc9457SAndroid Build Coastguard Worker     const float* input,
23*4bdc9457SAndroid Build Coastguard Worker     const float* weights,
24*4bdc9457SAndroid Build Coastguard Worker     const float* zero,
25*4bdc9457SAndroid Build Coastguard Worker     float* output,
26*4bdc9457SAndroid Build Coastguard Worker     uint32_t padding_top,
27*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28*4bdc9457SAndroid Build Coastguard Worker {
29*4bdc9457SAndroid Build Coastguard Worker   assert(input_height != 0);
30*4bdc9457SAndroid Build Coastguard Worker   assert(input_width != 0);
31*4bdc9457SAndroid Build Coastguard Worker   assert(input_width % sizeof(float) == 0);
32*4bdc9457SAndroid Build Coastguard Worker   assert(padding_top == 1);
33*4bdc9457SAndroid Build Coastguard Worker 
34*4bdc9457SAndroid Build Coastguard Worker   const v128_t vmask = wasm_v128_load(params->scalar.mask);
35*4bdc9457SAndroid Build Coastguard Worker   const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
36*4bdc9457SAndroid Build Coastguard Worker   const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
37*4bdc9457SAndroid Build Coastguard Worker 
38*4bdc9457SAndroid Build Coastguard Worker   const v128_t vw0123 = wasm_v128_load(weights);
39*4bdc9457SAndroid Build Coastguard Worker   const v128_t vw4567 = wasm_v128_load(weights + 4);
40*4bdc9457SAndroid Build Coastguard Worker   const v128_t vw89 = wasm_v128_load64_splat(weights + 8);
41*4bdc9457SAndroid Build Coastguard Worker 
42*4bdc9457SAndroid Build Coastguard Worker   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
43*4bdc9457SAndroid Build Coastguard Worker 
44*4bdc9457SAndroid Build Coastguard Worker   const float* i0 = zero;
45*4bdc9457SAndroid Build Coastguard Worker   const float* i1 = input;
46*4bdc9457SAndroid Build Coastguard Worker   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
47*4bdc9457SAndroid Build Coastguard Worker   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
48*4bdc9457SAndroid Build Coastguard Worker   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
49*4bdc9457SAndroid Build Coastguard Worker   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
50*4bdc9457SAndroid Build Coastguard Worker   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
51*4bdc9457SAndroid Build Coastguard Worker   const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
52*4bdc9457SAndroid Build Coastguard Worker 
53*4bdc9457SAndroid Build Coastguard Worker   float* o0 = output;
54*4bdc9457SAndroid Build Coastguard Worker   float* o1 = (float*) ((uintptr_t) o0 + input_width);
55*4bdc9457SAndroid Build Coastguard Worker   float* o2 = (float*) ((uintptr_t) o1 + input_width);
56*4bdc9457SAndroid Build Coastguard Worker   float* o3 = (float*) ((uintptr_t) o2 + input_width);
57*4bdc9457SAndroid Build Coastguard Worker   float* o4 = (float*) ((uintptr_t) o3 + input_width);
58*4bdc9457SAndroid Build Coastguard Worker   float* o5 = (float*) ((uintptr_t) o4 + input_width);
59*4bdc9457SAndroid Build Coastguard Worker 
60*4bdc9457SAndroid Build Coastguard Worker   size_t output_height = input_height;
61*4bdc9457SAndroid Build Coastguard Worker   do {
62*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 2) {
63*4bdc9457SAndroid Build Coastguard Worker       i2 = zero;
64*4bdc9457SAndroid Build Coastguard Worker       o1 = o0;
65*4bdc9457SAndroid Build Coastguard Worker     }
66*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 3) {
67*4bdc9457SAndroid Build Coastguard Worker       i3 = zero;
68*4bdc9457SAndroid Build Coastguard Worker       o2 = o1;
69*4bdc9457SAndroid Build Coastguard Worker     }
70*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 4) {
71*4bdc9457SAndroid Build Coastguard Worker       i4 = zero;
72*4bdc9457SAndroid Build Coastguard Worker       o3 = o2;
73*4bdc9457SAndroid Build Coastguard Worker     }
74*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 5) {
75*4bdc9457SAndroid Build Coastguard Worker       i5 = zero;
76*4bdc9457SAndroid Build Coastguard Worker       o4 = o3;
77*4bdc9457SAndroid Build Coastguard Worker     }
78*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 6) {
79*4bdc9457SAndroid Build Coastguard Worker       i6 = zero;
80*4bdc9457SAndroid Build Coastguard Worker       o5 = o4;
81*4bdc9457SAndroid Build Coastguard Worker     }
82*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(output_height < 7) {
83*4bdc9457SAndroid Build Coastguard Worker       i7 = zero;
84*4bdc9457SAndroid Build Coastguard Worker     }
85*4bdc9457SAndroid Build Coastguard Worker 
86*4bdc9457SAndroid Build Coastguard Worker     v128_t vi0x0123 = wasm_f32x4_const_splat(0.0f);
87*4bdc9457SAndroid Build Coastguard Worker     v128_t vi1x0123 = wasm_f32x4_const_splat(0.0f);
88*4bdc9457SAndroid Build Coastguard Worker     v128_t vi2x0123 = wasm_f32x4_const_splat(0.0f);
89*4bdc9457SAndroid Build Coastguard Worker     v128_t vi3x0123 = wasm_f32x4_const_splat(0.0f);
90*4bdc9457SAndroid Build Coastguard Worker     v128_t vi4x0123 = wasm_f32x4_const_splat(0.0f);
91*4bdc9457SAndroid Build Coastguard Worker     v128_t vi5x0123 = wasm_f32x4_const_splat(0.0f);
92*4bdc9457SAndroid Build Coastguard Worker     v128_t vi6x0123 = wasm_f32x4_const_splat(0.0f);
93*4bdc9457SAndroid Build Coastguard Worker     v128_t vi7x0123 = wasm_f32x4_const_splat(0.0f);
94*4bdc9457SAndroid Build Coastguard Worker 
95*4bdc9457SAndroid Build Coastguard Worker     v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
96*4bdc9457SAndroid Build Coastguard Worker     v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
97*4bdc9457SAndroid Build Coastguard Worker     v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
98*4bdc9457SAndroid Build Coastguard Worker     v128_t vi3x4567 = wasm_v128_load(i3); i3 += 4;
99*4bdc9457SAndroid Build Coastguard Worker     v128_t vi4x4567 = wasm_v128_load(i4); i4 += 4;
100*4bdc9457SAndroid Build Coastguard Worker     v128_t vi5x4567 = wasm_v128_load(i5); i5 += 4;
101*4bdc9457SAndroid Build Coastguard Worker     v128_t vi6x4567 = wasm_v128_load(i6); i6 += 4;
102*4bdc9457SAndroid Build Coastguard Worker     v128_t vi7x4567 = wasm_v128_load(i7); i7 += 4;
103*4bdc9457SAndroid Build Coastguard Worker 
104*4bdc9457SAndroid Build Coastguard Worker     size_t w = input_width;
105*4bdc9457SAndroid Build Coastguard Worker     for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
106*4bdc9457SAndroid Build Coastguard Worker       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
107*4bdc9457SAndroid Build Coastguard Worker       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
108*4bdc9457SAndroid Build Coastguard Worker       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
109*4bdc9457SAndroid Build Coastguard Worker       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
110*4bdc9457SAndroid Build Coastguard Worker       v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
111*4bdc9457SAndroid Build Coastguard Worker       v128_t vo5p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
112*4bdc9457SAndroid Build Coastguard Worker 
113*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
114*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
115*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
116*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi3x89AB = wasm_v128_load(i3); i3 += 4;
117*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi4x89AB = wasm_v128_load(i4); i4 += 4;
118*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi5x89AB = wasm_v128_load(i5); i5 += 4;
119*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi6x89AB = wasm_v128_load(i6); i6 += 4;
120*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi7x89AB = wasm_v128_load(i7); i7 += 4;
121*4bdc9457SAndroid Build Coastguard Worker 
122*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
123*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
124*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
125*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
126*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
127*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
128*4bdc9457SAndroid Build Coastguard Worker 
129*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
130*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
131*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
132*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
133*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
134*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
135*4bdc9457SAndroid Build Coastguard Worker 
136*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
137*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
138*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
139*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
140*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
141*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
142*4bdc9457SAndroid Build Coastguard Worker 
143*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
144*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
145*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
146*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
147*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
148*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
149*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
150*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
151*4bdc9457SAndroid Build Coastguard Worker 
152*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
153*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
154*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
155*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
156*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
157*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
158*4bdc9457SAndroid Build Coastguard Worker 
159*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
160*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
161*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
162*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
163*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
164*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
165*4bdc9457SAndroid Build Coastguard Worker 
166*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
167*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
168*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
169*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
170*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
171*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
172*4bdc9457SAndroid Build Coastguard Worker 
173*4bdc9457SAndroid Build Coastguard Worker       vi0x0123 = vi0x4567;
174*4bdc9457SAndroid Build Coastguard Worker       vi1x0123 = vi1x4567;
175*4bdc9457SAndroid Build Coastguard Worker       vi2x0123 = vi2x4567;
176*4bdc9457SAndroid Build Coastguard Worker       vi3x0123 = vi3x4567;
177*4bdc9457SAndroid Build Coastguard Worker       vi4x0123 = vi4x4567;
178*4bdc9457SAndroid Build Coastguard Worker       vi5x0123 = vi5x4567;
179*4bdc9457SAndroid Build Coastguard Worker       vi6x0123 = vi6x4567;
180*4bdc9457SAndroid Build Coastguard Worker       vi7x0123 = vi7x4567;
181*4bdc9457SAndroid Build Coastguard Worker 
182*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
183*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
184*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
185*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vi3x89AB, 1, 2, 3, 4);
186*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 1, 2, 3, 4);
187*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4);
188*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vi6x89AB, 1, 2, 3, 4);
189*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vi7x89AB, 1, 2, 3, 4);
190*4bdc9457SAndroid Build Coastguard Worker 
191*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
192*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
193*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
194*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
195*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
196*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
197*4bdc9457SAndroid Build Coastguard Worker 
198*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
199*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
200*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
201*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
202*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
203*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
204*4bdc9457SAndroid Build Coastguard Worker 
205*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
206*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
207*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
208*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
209*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
210*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
211*4bdc9457SAndroid Build Coastguard Worker 
212*4bdc9457SAndroid Build Coastguard Worker       vi0x4567 = vi0x89AB;
213*4bdc9457SAndroid Build Coastguard Worker       vi1x4567 = vi1x89AB;
214*4bdc9457SAndroid Build Coastguard Worker       vi2x4567 = vi2x89AB;
215*4bdc9457SAndroid Build Coastguard Worker       vi3x4567 = vi3x89AB;
216*4bdc9457SAndroid Build Coastguard Worker       vi4x4567 = vi4x89AB;
217*4bdc9457SAndroid Build Coastguard Worker       vi5x4567 = vi5x89AB;
218*4bdc9457SAndroid Build Coastguard Worker       vi6x4567 = vi6x89AB;
219*4bdc9457SAndroid Build Coastguard Worker       vi7x4567 = vi7x89AB;
220*4bdc9457SAndroid Build Coastguard Worker 
221*4bdc9457SAndroid Build Coastguard Worker 
222*4bdc9457SAndroid Build Coastguard Worker       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
223*4bdc9457SAndroid Build Coastguard Worker       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
224*4bdc9457SAndroid Build Coastguard Worker       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
225*4bdc9457SAndroid Build Coastguard Worker       v128_t vo3 = wasm_f32x4_max(vo3p0, vmin);
226*4bdc9457SAndroid Build Coastguard Worker       v128_t vo4 = wasm_f32x4_max(vo4p0, vmin);
227*4bdc9457SAndroid Build Coastguard Worker       v128_t vo5 = wasm_f32x4_max(vo5p0, vmin);
228*4bdc9457SAndroid Build Coastguard Worker       vo0 = wasm_f32x4_min(vo0, vmax);
229*4bdc9457SAndroid Build Coastguard Worker       vo1 = wasm_f32x4_min(vo1, vmax);
230*4bdc9457SAndroid Build Coastguard Worker       vo2 = wasm_f32x4_min(vo2, vmax);
231*4bdc9457SAndroid Build Coastguard Worker       vo3 = wasm_f32x4_min(vo3, vmax);
232*4bdc9457SAndroid Build Coastguard Worker       vo4 = wasm_f32x4_min(vo4, vmax);
233*4bdc9457SAndroid Build Coastguard Worker       vo5 = wasm_f32x4_min(vo5, vmax);
234*4bdc9457SAndroid Build Coastguard Worker 
235*4bdc9457SAndroid Build Coastguard Worker       wasm_v128_store(o5, vo5); o5 += 4;
236*4bdc9457SAndroid Build Coastguard Worker       wasm_v128_store(o4, vo4); o4 += 4;
237*4bdc9457SAndroid Build Coastguard Worker       wasm_v128_store(o3, vo3); o3 += 4;
238*4bdc9457SAndroid Build Coastguard Worker       wasm_v128_store(o2, vo2); o2 += 4;
239*4bdc9457SAndroid Build Coastguard Worker       wasm_v128_store(o1, vo1); o1 += 4;
240*4bdc9457SAndroid Build Coastguard Worker       wasm_v128_store(o0, vo0); o0 += 4;
241*4bdc9457SAndroid Build Coastguard Worker     }
242*4bdc9457SAndroid Build Coastguard Worker     // Always process the last block of 1..4 pixels.
243*4bdc9457SAndroid Build Coastguard Worker     assert(w >= 1 * sizeof(float));
244*4bdc9457SAndroid Build Coastguard Worker     assert(w <= 4 * sizeof(float));
245*4bdc9457SAndroid Build Coastguard Worker     {
246*4bdc9457SAndroid Build Coastguard Worker       v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
247*4bdc9457SAndroid Build Coastguard Worker       v128_t vo1p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
248*4bdc9457SAndroid Build Coastguard Worker       v128_t vo2p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
249*4bdc9457SAndroid Build Coastguard Worker       v128_t vo3p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
250*4bdc9457SAndroid Build Coastguard Worker       v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
251*4bdc9457SAndroid Build Coastguard Worker       v128_t vo5p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
252*4bdc9457SAndroid Build Coastguard Worker 
253*4bdc9457SAndroid Build Coastguard Worker       vi0x4567 = wasm_v128_and(vmask, vi0x4567);
254*4bdc9457SAndroid Build Coastguard Worker       vi1x4567 = wasm_v128_and(vmask, vi1x4567);
255*4bdc9457SAndroid Build Coastguard Worker       vi2x4567 = wasm_v128_and(vmask, vi2x4567);
256*4bdc9457SAndroid Build Coastguard Worker       vi3x4567 = wasm_v128_and(vmask, vi3x4567);
257*4bdc9457SAndroid Build Coastguard Worker       vi4x4567 = wasm_v128_and(vmask, vi4x4567);
258*4bdc9457SAndroid Build Coastguard Worker       vi5x4567 = wasm_v128_and(vmask, vi5x4567);
259*4bdc9457SAndroid Build Coastguard Worker       vi6x4567 = wasm_v128_and(vmask, vi6x4567);
260*4bdc9457SAndroid Build Coastguard Worker       vi7x4567 = wasm_v128_and(vmask, vi7x4567);
261*4bdc9457SAndroid Build Coastguard Worker 
262*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
263*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
264*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
265*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
266*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
267*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
268*4bdc9457SAndroid Build Coastguard Worker 
269*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
270*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
271*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
272*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
273*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
274*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)));
275*4bdc9457SAndroid Build Coastguard Worker 
276*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
277*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
278*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
279*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
280*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
281*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)));
282*4bdc9457SAndroid Build Coastguard Worker 
283*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
284*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
285*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
286*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi3x3456 = wasm_v32x4_shuffle(vi3x0123, vi3x4567, 3, 4, 5, 6);
287*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi4x3456 = wasm_v32x4_shuffle(vi4x0123, vi4x4567, 3, 4, 5, 6);
288*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6);
289*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi6x3456 = wasm_v32x4_shuffle(vi6x0123, vi6x4567, 3, 4, 5, 6);
290*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6);
291*4bdc9457SAndroid Build Coastguard Worker 
292*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
293*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
294*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
295*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
296*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
297*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)));
298*4bdc9457SAndroid Build Coastguard Worker 
299*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
300*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
301*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
302*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
303*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
304*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
305*4bdc9457SAndroid Build Coastguard Worker 
306*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
307*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
308*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
309*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
310*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
311*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
312*4bdc9457SAndroid Build Coastguard Worker 
313*4bdc9457SAndroid Build Coastguard Worker       const v128_t vzero = wasm_f32x4_const_splat(0.0f);
314*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
315*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
316*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
317*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi3x5678 = wasm_v32x4_shuffle(vi3x4567, vzero, 1, 2, 3, 4);
318*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi4x5678 = wasm_v32x4_shuffle(vi4x4567, vzero, 1, 2, 3, 4);
319*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4);
320*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi6x5678 = wasm_v32x4_shuffle(vi6x4567, vzero, 1, 2, 3, 4);
321*4bdc9457SAndroid Build Coastguard Worker       const v128_t vi7x5678 = wasm_v32x4_shuffle(vi7x4567, vzero, 1, 2, 3, 4);
322*4bdc9457SAndroid Build Coastguard Worker 
323*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
324*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
325*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
326*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
327*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
328*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
329*4bdc9457SAndroid Build Coastguard Worker 
330*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
331*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
332*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
333*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
334*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
335*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
336*4bdc9457SAndroid Build Coastguard Worker 
337*4bdc9457SAndroid Build Coastguard Worker       vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
338*4bdc9457SAndroid Build Coastguard Worker       vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
339*4bdc9457SAndroid Build Coastguard Worker       vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
340*4bdc9457SAndroid Build Coastguard Worker       vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
341*4bdc9457SAndroid Build Coastguard Worker       vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
342*4bdc9457SAndroid Build Coastguard Worker       vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi7x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
343*4bdc9457SAndroid Build Coastguard Worker 
344*4bdc9457SAndroid Build Coastguard Worker 
345*4bdc9457SAndroid Build Coastguard Worker       v128_t vo0 = wasm_f32x4_max(vo0p0, vmin);
346*4bdc9457SAndroid Build Coastguard Worker       v128_t vo1 = wasm_f32x4_max(vo1p0, vmin);
347*4bdc9457SAndroid Build Coastguard Worker       v128_t vo2 = wasm_f32x4_max(vo2p0, vmin);
348*4bdc9457SAndroid Build Coastguard Worker       v128_t vo3 = wasm_f32x4_max(vo3p0, vmin);
349*4bdc9457SAndroid Build Coastguard Worker       v128_t vo4 = wasm_f32x4_max(vo4p0, vmin);
350*4bdc9457SAndroid Build Coastguard Worker       v128_t vo5 = wasm_f32x4_max(vo5p0, vmin);
351*4bdc9457SAndroid Build Coastguard Worker       vo0 = wasm_f32x4_min(vo0, vmax);
352*4bdc9457SAndroid Build Coastguard Worker       vo1 = wasm_f32x4_min(vo1, vmax);
353*4bdc9457SAndroid Build Coastguard Worker       vo2 = wasm_f32x4_min(vo2, vmax);
354*4bdc9457SAndroid Build Coastguard Worker       vo3 = wasm_f32x4_min(vo3, vmax);
355*4bdc9457SAndroid Build Coastguard Worker       vo4 = wasm_f32x4_min(vo4, vmax);
356*4bdc9457SAndroid Build Coastguard Worker       vo5 = wasm_f32x4_min(vo5, vmax);
357*4bdc9457SAndroid Build Coastguard Worker 
358*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(w == 4 * sizeof(float)) {
359*4bdc9457SAndroid Build Coastguard Worker         wasm_v128_store(o5, vo5); o5 += 4;
360*4bdc9457SAndroid Build Coastguard Worker         wasm_v128_store(o4, vo4); o4 += 4;
361*4bdc9457SAndroid Build Coastguard Worker         wasm_v128_store(o3, vo3); o3 += 4;
362*4bdc9457SAndroid Build Coastguard Worker         wasm_v128_store(o2, vo2); o2 += 4;
363*4bdc9457SAndroid Build Coastguard Worker         wasm_v128_store(o1, vo1); o1 += 4;
364*4bdc9457SAndroid Build Coastguard Worker         wasm_v128_store(o0, vo0); o0 += 4;
365*4bdc9457SAndroid Build Coastguard Worker       } else {
366*4bdc9457SAndroid Build Coastguard Worker         if (w & (2 * sizeof(float))) {
367*4bdc9457SAndroid Build Coastguard Worker           *((double*) o5) = wasm_f64x2_extract_lane(vo5, 0); o5 += 2;
368*4bdc9457SAndroid Build Coastguard Worker           *((double*) o4) = wasm_f64x2_extract_lane(vo4, 0); o4 += 2;
369*4bdc9457SAndroid Build Coastguard Worker           *((double*) o3) = wasm_f64x2_extract_lane(vo3, 0); o3 += 2;
370*4bdc9457SAndroid Build Coastguard Worker           *((double*) o2) = wasm_f64x2_extract_lane(vo2, 0); o2 += 2;
371*4bdc9457SAndroid Build Coastguard Worker           *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2;
372*4bdc9457SAndroid Build Coastguard Worker           *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
373*4bdc9457SAndroid Build Coastguard Worker 
374*4bdc9457SAndroid Build Coastguard Worker           vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
375*4bdc9457SAndroid Build Coastguard Worker           vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1);
376*4bdc9457SAndroid Build Coastguard Worker           vo2 = wasm_v32x4_shuffle(vo2, vo2, 2, 3, 0, 1);
377*4bdc9457SAndroid Build Coastguard Worker           vo3 = wasm_v32x4_shuffle(vo3, vo3, 2, 3, 0, 1);
378*4bdc9457SAndroid Build Coastguard Worker           vo4 = wasm_v32x4_shuffle(vo4, vo4, 2, 3, 0, 1);
379*4bdc9457SAndroid Build Coastguard Worker           vo5 = wasm_v32x4_shuffle(vo5, vo5, 2, 3, 0, 1);
380*4bdc9457SAndroid Build Coastguard Worker         }
381*4bdc9457SAndroid Build Coastguard Worker         if (w & (1 * sizeof(float))) {
382*4bdc9457SAndroid Build Coastguard Worker           *o5 = wasm_f32x4_extract_lane(vo5, 0); o5 += 1;
383*4bdc9457SAndroid Build Coastguard Worker           *o4 = wasm_f32x4_extract_lane(vo4, 0); o4 += 1;
384*4bdc9457SAndroid Build Coastguard Worker           *o3 = wasm_f32x4_extract_lane(vo3, 0); o3 += 1;
385*4bdc9457SAndroid Build Coastguard Worker           *o2 = wasm_f32x4_extract_lane(vo2, 0); o2 += 1;
386*4bdc9457SAndroid Build Coastguard Worker           *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1;
387*4bdc9457SAndroid Build Coastguard Worker           *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
388*4bdc9457SAndroid Build Coastguard Worker         }
389*4bdc9457SAndroid Build Coastguard Worker       }
390*4bdc9457SAndroid Build Coastguard Worker     }
391*4bdc9457SAndroid Build Coastguard Worker 
392*4bdc9457SAndroid Build Coastguard Worker     i0 = (const float*) ((uintptr_t) i6 - input_decrement);
393*4bdc9457SAndroid Build Coastguard Worker     i1 = (const float*) ((uintptr_t) i7 - input_decrement);
394*4bdc9457SAndroid Build Coastguard Worker     i2 = (const float*) ((uintptr_t) i1 + input_width);
395*4bdc9457SAndroid Build Coastguard Worker     i3 = (const float*) ((uintptr_t) i2 + input_width);
396*4bdc9457SAndroid Build Coastguard Worker     i4 = (const float*) ((uintptr_t) i3 + input_width);
397*4bdc9457SAndroid Build Coastguard Worker     i5 = (const float*) ((uintptr_t) i4 + input_width);
398*4bdc9457SAndroid Build Coastguard Worker     i6 = (const float*) ((uintptr_t) i5 + input_width);
399*4bdc9457SAndroid Build Coastguard Worker     i7 = (const float*) ((uintptr_t) i6 + input_width);
400*4bdc9457SAndroid Build Coastguard Worker 
401*4bdc9457SAndroid Build Coastguard Worker     o0 = o5;
402*4bdc9457SAndroid Build Coastguard Worker     o1 = (float*) ((uintptr_t) o0 + input_width);
403*4bdc9457SAndroid Build Coastguard Worker     o2 = (float*) ((uintptr_t) o1 + input_width);
404*4bdc9457SAndroid Build Coastguard Worker     o3 = (float*) ((uintptr_t) o2 + input_width);
405*4bdc9457SAndroid Build Coastguard Worker     o4 = (float*) ((uintptr_t) o3 + input_width);
406*4bdc9457SAndroid Build Coastguard Worker     o5 = (float*) ((uintptr_t) o4 + input_width);
407*4bdc9457SAndroid Build Coastguard Worker 
408*4bdc9457SAndroid Build Coastguard Worker     output_height = doz(output_height, 6);
409*4bdc9457SAndroid Build Coastguard Worker   } while (output_height != 0);
410*4bdc9457SAndroid Build Coastguard Worker }
411