1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/3x3p1-wasmsimd-splat.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 #include <xnnpack/dwconv.h>
15 #include <xnnpack/math.h>
16
17
18
xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4(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 void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4(
20 size_t input_height,
21 size_t input_width,
22 const float* input,
23 const float* weights,
24 const float* zero,
25 float* output,
26 uint32_t padding_top,
27 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(input_height != 0);
30 assert(input_width != 0);
31 assert(input_width % sizeof(float) == 0);
32 assert(padding_top == 1);
33
34 const v128_t vmask = wasm_v128_load(params->scalar.mask);
35 const v128_t vmax = wasm_v128_load32_splat(¶ms->scalar.max);
36 const v128_t vmin = wasm_v128_load32_splat(¶ms->scalar.min);
37
38 const v128_t vw0123 = wasm_v128_load(weights);
39 const v128_t vw4567 = wasm_v128_load(weights + 4);
40 const v128_t vw89 = wasm_v128_load64_splat(weights + 8);
41
42 const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
43
44 const float* i0 = zero;
45 const float* i1 = input;
46 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
47
48 float* o0 = output;
49
50 size_t output_height = input_height;
51 do {
52 if XNN_UNPREDICTABLE(output_height < 2) {
53 i2 = zero;
54 }
55
56 v128_t vi0x0123 = wasm_f32x4_const_splat(0.0f);
57 v128_t vi1x0123 = wasm_f32x4_const_splat(0.0f);
58 v128_t vi2x0123 = wasm_f32x4_const_splat(0.0f);
59
60 v128_t vi0x4567 = wasm_v128_load(i0); i0 += 4;
61 v128_t vi1x4567 = wasm_v128_load(i1); i1 += 4;
62 v128_t vi2x4567 = wasm_v128_load(i2); i2 += 4;
63
64 size_t w = input_width;
65 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
66 v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
67
68 const v128_t vi0x89AB = wasm_v128_load(i0); i0 += 4;
69 const v128_t vi1x89AB = wasm_v128_load(i1); i1 += 4;
70 const v128_t vi2x89AB = wasm_v128_load(i2); i2 += 4;
71
72 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
73
74 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1));
75
76 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0));
77
78 const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
79 const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
80 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
81
82 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1));
83
84 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
85
86 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
87
88 vi0x0123 = vi0x4567;
89 vi1x0123 = vi1x4567;
90 vi2x0123 = vi2x4567;
91
92 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4);
93 const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vi1x89AB, 1, 2, 3, 4);
94 const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vi2x89AB, 1, 2, 3, 4);
95
96 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
97
98 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
99
100 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
101
102 vi0x4567 = vi0x89AB;
103 vi1x4567 = vi1x89AB;
104 vi2x4567 = vi2x89AB;
105
106 vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
107 vo0p2 = wasm_f32x4_add(vo0p2, vo0p3);
108 vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
109
110 v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
111 vo0 = wasm_f32x4_pmin(vmax, vo0);
112
113 wasm_v128_store(o0, vo0); o0 += 4;
114 }
115 // Always process the last block of 1..4 pixels.
116 assert(w >= 1 * sizeof(float));
117 assert(w <= 4 * sizeof(float));
118 {
119 v128_t vo0p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
120
121 vi0x4567 = wasm_v128_and(vmask, vi0x4567);
122 vi1x4567 = wasm_v128_and(vmask, vi1x4567);
123 vi2x4567 = wasm_v128_and(vmask, vi2x4567);
124
125 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x4567, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)));
126
127 v128_t vo0p1 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1));
128
129 v128_t vo0p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0));
130
131 const v128_t vi0x3456 = wasm_v32x4_shuffle(vi0x0123, vi0x4567, 3, 4, 5, 6);
132 const v128_t vi1x3456 = wasm_v32x4_shuffle(vi1x0123, vi1x4567, 3, 4, 5, 6);
133 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6);
134
135 v128_t vo0p3 = wasm_f32x4_mul(vi0x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1));
136
137 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)));
138
139 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3)));
140
141 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
142 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4);
143 const v128_t vi1x5678 = wasm_v32x4_shuffle(vi1x4567, vzero, 1, 2, 3, 4);
144 const v128_t vi2x5678 = wasm_v32x4_shuffle(vi2x4567, vzero, 1, 2, 3, 4);
145
146 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3)));
147
148 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2)));
149
150 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1)));
151
152 vo0p0 = wasm_f32x4_add(vo0p0, vo0p1);
153 vo0p2 = wasm_f32x4_add(vo0p2, vo0p3);
154 vo0p0 = wasm_f32x4_add(vo0p0, vo0p2);
155
156 v128_t vo0 = wasm_f32x4_pmax(vmin, vo0p0);
157 vo0 = wasm_f32x4_pmin(vmax, vo0);
158
159 if XNN_LIKELY(w == 4 * sizeof(float)) {
160 wasm_v128_store(o0, vo0); o0 += 4;
161 } else {
162 if (w & (2 * sizeof(float))) {
163 *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2;
164
165 vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1);
166 }
167 if (w & (1 * sizeof(float))) {
168 *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1;
169 }
170 }
171 }
172
173 i0 = (const float*) ((uintptr_t) i1 - input_decrement);
174 i1 = (const float*) ((uintptr_t) i2 - input_decrement);
175 i2 = (const float*) ((uintptr_t) i1 + input_width);
176
177
178 } while (--output_height != 0);
179 }
180