1 // Copyright 2019 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 <psimd.h>
9
10 #include <xnnpack/argmaxpool.h>
11
12
xnn_f32_argmaxpool_ukernel_4x__psimd_c4(size_t output_pixels,size_t pooling_elements,size_t channels,const float ** input,size_t input_offset,float * output,uint32_t * index,size_t input_increment,size_t output_increment,const union xnn_f32_output_params params[restrict static1])13 void xnn_f32_argmaxpool_ukernel_4x__psimd_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,
21 size_t input_increment,
22 size_t output_increment,
23 const union xnn_f32_output_params params[restrict static 1])
24 {
25 assert(output_pixels != 0);
26 assert(pooling_elements != 0);
27 assert(pooling_elements <= 4);
28 assert(channels != 0);
29
30 const psimd_f32 voutput_max = psimd_load_splat_f32(¶ms->scalar.max);
31 const psimd_f32 voutput_min = psimd_load_splat_f32(¶ms->scalar.min);
32 do {
33 const float* i0 = input[0];
34 const float* i1 = input[1];
35 const float* i2 = input[2];
36 const float* i3 = input[3];
37 i0 = (const float*) ((uintptr_t) i0 + input_offset);
38 i1 = (const float*) ((uintptr_t) i1 + input_offset);
39 i2 = (const float*) ((uintptr_t) i2 + input_offset);
40 i3 = (const float*) ((uintptr_t) i3 + input_offset);
41 if (pooling_elements < 2) {
42 i1 = i0;
43 }
44 if (pooling_elements <= 2) {
45 i2 = i0;
46 }
47 if (pooling_elements != 4) {
48 i3 = i0;
49 }
50
51 size_t c = channels;
52 for (; c >= 4; c -= 4) {
53 const psimd_f32 vi0 = psimd_load_f32(i0);
54 i0 += 4;
55 const psimd_f32 vi1 = psimd_load_f32(i1);
56 i1 += 4;
57 const psimd_f32 vi2 = psimd_load_f32(i2);
58 i2 += 4;
59 const psimd_f32 vi3 = psimd_load_f32(i3);
60 i3 += 4;
61
62 psimd_f32 vmax = vi0;
63 psimd_u32 vidx = psimd_splat_u32(0);
64
65 const psimd_s32 vm1 = vi1 > vmax;
66 vmax = psimd_blend_f32(vm1, vi1, vmax);
67 vidx = psimd_blend_u32(vm1, psimd_splat_u32(1), vidx);
68
69 const psimd_s32 vm2 = vi2 > vmax;
70 vmax = psimd_blend_f32(vm2, vi2, vmax);
71 vidx = psimd_blend_u32(vm2, psimd_splat_u32(2), vidx);
72
73 const psimd_s32 vm3 = vi3 > vmax;
74 vmax = psimd_blend_f32(vm3, vi3, vmax);
75 vidx = psimd_blend_u32(vm3, psimd_splat_u32(3), vidx);
76
77 const psimd_f32 vout = psimd_max_f32(psimd_min_f32(vmax, voutput_max), voutput_min);
78
79 psimd_store_f32(output, vout);
80 output += 4;
81 psimd_store_u32(index, vidx);
82 index += 4;
83 }
84 if (c != 0) {
85 const psimd_f32 vi0 = psimd_load_f32(i0);
86 const psimd_f32 vi1 = psimd_load_f32(i1);
87 const psimd_f32 vi2 = psimd_load_f32(i2);
88 const psimd_f32 vi3 = psimd_load_f32(i3);
89
90 psimd_f32 vmax = vi0;
91 psimd_u32 vidx = psimd_splat_u32(0);
92
93 const psimd_s32 vm1 = vi1 > vmax;
94 vmax = psimd_blend_f32(vm1, vi1, vmax);
95 vidx = psimd_blend_u32(vm1, psimd_splat_u32(1), vidx);
96
97 const psimd_s32 vm2 = vi2 > vmax;
98 vmax = psimd_blend_f32(vm2, vi2, vmax);
99 vidx = psimd_blend_u32(vm2, psimd_splat_u32(2), vidx);
100
101 const psimd_s32 vm3 = vi3 > vmax;
102 vmax = psimd_blend_f32(vm3, vi3, vmax);
103 vidx = psimd_blend_u32(vm3, psimd_splat_u32(3), vidx);
104
105 psimd_f32 vout = psimd_max_f32(psimd_min_f32(vmax, voutput_max), voutput_min);
106
107 if (c & 2) {
108 psimd_store2_f32(output, vout);
109 psimd_store2_u32(index, vidx);
110 vout = psimd_concat_hi_f32(vout, vout);
111 vidx = psimd_concat_hi_u32(vidx, vidx);
112 output += 2;
113 index += 2;
114 }
115 if (c & 1) {
116 psimd_store1_f32(output, vout);
117 psimd_store1_u32(index, vidx);
118 output += 1;
119 index += 1;
120 }
121 }
122 input = (const float**) ((uintptr_t) input + input_increment);
123 output = (float*) ((uintptr_t) output + output_increment);
124 } while (--output_pixels != 0);
125 }
126