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 #include <assert.h>
7
8 #include <wasm_simd128.h>
9
10 #include <xnnpack/argmaxpool.h>
11
12
xnn_f32_argmaxpool_ukernel_4x__wasmsimd_c4(size_t output_pixels,size_t pooling_elements,size_t channels,const float ** input,size_t input_offset,float * output,uint32_t * index_ptr,size_t input_increment,size_t output_increment)13 void xnn_f32_argmaxpool_ukernel_4x__wasmsimd_c4(
14 size_t output_pixels,
15 size_t pooling_elements,
16 size_t channels,
17 const float** input,
18 size_t input_offset,
19 float* output,
20 uint32_t* index_ptr,
21 size_t input_increment,
22 size_t output_increment) XNN_OOB_READS
23 {
24 assert(output_pixels != 0);
25 assert(pooling_elements != 0);
26 assert(pooling_elements <= 4);
27 assert(channels != 0);
28
29 float* index = (float*) index_ptr;
30 do {
31 const float* i0 = input[0];
32 const float* i1 = input[1];
33 const float* i2 = input[2];
34 const float* i3 = input[3];
35 i0 = (const float*) ((uintptr_t) i0 + input_offset);
36 i1 = (const float*) ((uintptr_t) i1 + input_offset);
37 i2 = (const float*) ((uintptr_t) i2 + input_offset);
38 i3 = (const float*) ((uintptr_t) i3 + input_offset);
39 if (pooling_elements < 2) {
40 i1 = i0;
41 }
42 if (pooling_elements <= 2) {
43 i2 = i0;
44 }
45 if (pooling_elements != 4) {
46 i3 = i0;
47 }
48
49 size_t c = channels;
50 for (; c >= 4; c -= 4) {
51 const v128_t vi0 = wasm_v128_load(i0);
52 i0 += 4;
53 const v128_t vi1 = wasm_v128_load(i1);
54 i1 += 4;
55 const v128_t vi2 = wasm_v128_load(i2);
56 i2 += 4;
57 const v128_t vi3 = wasm_v128_load(i3);
58 i3 += 4;
59
60 v128_t vmax = vi0;
61 v128_t vidx = wasm_i32x4_const_splat(0);
62
63 const v128_t vm1 = wasm_f32x4_gt(vi1, vmax);
64 vmax = wasm_v128_bitselect(vi1, vmax, vm1);
65 vidx = wasm_v128_bitselect(wasm_i32x4_const_splat(1), vidx, vm1);
66
67 const v128_t vm2 = wasm_f32x4_gt(vi2, vmax);
68 vmax = wasm_v128_bitselect(vi2, vmax, vm2);
69 vidx = wasm_v128_bitselect(wasm_i32x4_const_splat(2), vidx, vm2);
70
71 const v128_t vm3 = wasm_f32x4_gt(vi3, vmax);
72 vmax = wasm_v128_bitselect(vi3, vmax, vm3);
73 vidx = wasm_v128_bitselect(wasm_i32x4_const_splat(3), vidx, vm3);
74
75 wasm_v128_store(output, vmax);
76 output += 4;
77 wasm_v128_store(index, vidx);
78 index += 4;
79 }
80 if (c != 0) {
81 const v128_t vi0 = wasm_v128_load(i0);
82 const v128_t vi1 = wasm_v128_load(i1);
83 const v128_t vi2 = wasm_v128_load(i2);
84 const v128_t vi3 = wasm_v128_load(i3);
85
86 v128_t vmax = vi0;
87 v128_t vidx = wasm_i32x4_const_splat(0);
88
89 const v128_t vm1 = wasm_f32x4_gt(vi1, vmax);
90 vmax = wasm_v128_bitselect(vi1, vmax, vm1);
91 vidx = wasm_v128_bitselect(wasm_i32x4_const_splat(1), vidx, vm1);
92
93 const v128_t vm2 = wasm_f32x4_gt(vi2, vmax);
94 vmax = wasm_v128_bitselect(vi2, vmax, vm2);
95 vidx = wasm_v128_bitselect(wasm_i32x4_const_splat(2), vidx, vm2);
96
97 const v128_t vm3 = wasm_f32x4_gt(vi3, vmax);
98 vmax = wasm_v128_bitselect(vi3, vmax, vm3);
99 vidx = wasm_v128_bitselect(wasm_i32x4_const_splat(3), vidx, vm3);
100
101 if (c & 2) {
102 *((double*) output) = wasm_f64x2_extract_lane(vmax, 0);
103 *((double*) index) = wasm_f64x2_extract_lane(vidx, 0);
104 vmax = wasm_v32x4_shuffle(vmax, vmax, 2, 3, 2, 3);
105 vidx = wasm_v32x4_shuffle(vidx, vidx, 2, 3, 2, 3);
106 output += 2;
107 index += 2;
108 }
109 if (c & 1) {
110 *output++ = wasm_f32x4_extract_lane(vmax, 0);
111 *index++ = wasm_f32x4_extract_lane(vidx, 0);
112 }
113 }
114 input = (const float**) ((uintptr_t) input + input_increment);
115 output = (float*) ((uintptr_t) output + output_increment);
116 } while (--output_pixels != 0);
117 }
118