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_9x__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_9x__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 <= 9);
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 const float* i4 = input[4];
38 const float* i5 = input[5];
39 const float* i6 = input[6];
40 const float* i7 = input[7];
41 const float* i8 = input[8];
42 i0 = (const float*) ((uintptr_t) i0 + input_offset);
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 i2 = (const float*) ((uintptr_t) i2 + input_offset);
45 i3 = (const float*) ((uintptr_t) i3 + input_offset);
46 i4 = (const float*) ((uintptr_t) i4 + input_offset);
47 i5 = (const float*) ((uintptr_t) i5 + input_offset);
48 i6 = (const float*) ((uintptr_t) i6 + input_offset);
49 i7 = (const float*) ((uintptr_t) i7 + input_offset);
50 i8 = (const float*) ((uintptr_t) i8 + input_offset);
51 if (pooling_elements < 2) {
52 i1 = i0;
53 }
54 if (pooling_elements <= 2) {
55 i2 = i0;
56 }
57 if (pooling_elements < 4) {
58 i3 = i0;
59 }
60 if (pooling_elements <= 4) {
61 i4 = i0;
62 }
63 if (pooling_elements < 6) {
64 i5 = i0;
65 }
66 if (pooling_elements <= 6) {
67 i6 = i0;
68 }
69 if (pooling_elements < 8) {
70 i7 = i0;
71 }
72 if (pooling_elements <= 8) {
73 i8 = i0;
74 }
75
76 size_t c = channels;
77 for (; c >= 4; c -= 4) {
78 const psimd_f32 vi0 = psimd_load_f32(i0);
79 i0 += 4;
80 const psimd_f32 vi1 = psimd_load_f32(i1);
81 i1 += 4;
82 const psimd_f32 vi2 = psimd_load_f32(i2);
83 i2 += 4;
84 const psimd_f32 vi3 = psimd_load_f32(i3);
85 i3 += 4;
86 const psimd_f32 vi4 = psimd_load_f32(i4);
87 i4 += 4;
88 const psimd_f32 vi5 = psimd_load_f32(i5);
89 i5 += 4;
90 const psimd_f32 vi6 = psimd_load_f32(i6);
91 i6 += 4;
92 const psimd_f32 vi7 = psimd_load_f32(i7);
93 i7 += 4;
94 const psimd_f32 vi8 = psimd_load_f32(i8);
95 i8 += 4;
96
97 psimd_f32 vmax = vi0;
98 psimd_u32 vidx = psimd_splat_u32(0);
99
100 const psimd_s32 vm1 = vi1 > vmax;
101 vmax = psimd_blend_f32(vm1, vi1, vmax);
102 vidx = psimd_blend_u32(vm1, psimd_splat_u32(1), vidx);
103
104 const psimd_s32 vm2 = vi2 > vmax;
105 vmax = psimd_blend_f32(vm2, vi2, vmax);
106 vidx = psimd_blend_u32(vm2, psimd_splat_u32(2), vidx);
107
108 const psimd_s32 vm3 = vi3 > vmax;
109 vmax = psimd_blend_f32(vm3, vi3, vmax);
110 vidx = psimd_blend_u32(vm3, psimd_splat_u32(3), vidx);
111
112 const psimd_s32 vm4 = vi4 > vmax;
113 vmax = psimd_blend_f32(vm4, vi4, vmax);
114 vidx = psimd_blend_u32(vm4, psimd_splat_u32(4), vidx);
115
116 const psimd_s32 vm5 = vi5 > vmax;
117 vmax = psimd_blend_f32(vm5, vi5, vmax);
118 vidx = psimd_blend_u32(vm5, psimd_splat_u32(5), vidx);
119
120 const psimd_s32 vm6 = vi6 > vmax;
121 vmax = psimd_blend_f32(vm6, vi6, vmax);
122 vidx = psimd_blend_u32(vm6, psimd_splat_u32(6), vidx);
123
124 const psimd_s32 vm7 = vi7 > vmax;
125 vmax = psimd_blend_f32(vm7, vi7, vmax);
126 vidx = psimd_blend_u32(vm7, psimd_splat_u32(7), vidx);
127
128 const psimd_s32 vm8 = vi8 > vmax;
129 vmax = psimd_blend_f32(vm8, vi8, vmax);
130 vidx = psimd_blend_u32(vm8, psimd_splat_u32(8), vidx);
131
132 const psimd_f32 vout = psimd_max_f32(psimd_min_f32(vmax, voutput_max), voutput_min);
133
134 psimd_store_f32(output, vout);
135 output += 4;
136 psimd_store_u32(index, vidx);
137 index += 4;
138 }
139 if (c != 0) {
140 const psimd_f32 vi0 = psimd_load_f32(i0);
141 const psimd_f32 vi1 = psimd_load_f32(i1);
142 const psimd_f32 vi2 = psimd_load_f32(i2);
143 const psimd_f32 vi3 = psimd_load_f32(i3);
144 const psimd_f32 vi4 = psimd_load_f32(i4);
145 const psimd_f32 vi5 = psimd_load_f32(i5);
146 const psimd_f32 vi6 = psimd_load_f32(i6);
147 const psimd_f32 vi7 = psimd_load_f32(i7);
148 const psimd_f32 vi8 = psimd_load_f32(i8);
149
150 psimd_f32 vmax = vi0;
151 psimd_u32 vidx = psimd_splat_u32(0);
152
153 const psimd_s32 vm1 = vi1 > vmax;
154 vmax = psimd_blend_f32(vm1, vi1, vmax);
155 vidx = psimd_blend_u32(vm1, psimd_splat_u32(1), vidx);
156
157 const psimd_s32 vm2 = vi2 > vmax;
158 vmax = psimd_blend_f32(vm2, vi2, vmax);
159 vidx = psimd_blend_u32(vm2, psimd_splat_u32(2), vidx);
160
161 const psimd_s32 vm3 = vi3 > vmax;
162 vmax = psimd_blend_f32(vm3, vi3, vmax);
163 vidx = psimd_blend_u32(vm3, psimd_splat_u32(3), vidx);
164
165 const psimd_s32 vm4 = vi4 > vmax;
166 vmax = psimd_blend_f32(vm4, vi4, vmax);
167 vidx = psimd_blend_u32(vm4, psimd_splat_u32(4), vidx);
168
169 const psimd_s32 vm5 = vi5 > vmax;
170 vmax = psimd_blend_f32(vm5, vi5, vmax);
171 vidx = psimd_blend_u32(vm5, psimd_splat_u32(5), vidx);
172
173 const psimd_s32 vm6 = vi6 > vmax;
174 vmax = psimd_blend_f32(vm6, vi6, vmax);
175 vidx = psimd_blend_u32(vm6, psimd_splat_u32(6), vidx);
176
177 const psimd_s32 vm7 = vi7 > vmax;
178 vmax = psimd_blend_f32(vm7, vi7, vmax);
179 vidx = psimd_blend_u32(vm7, psimd_splat_u32(7), vidx);
180
181 const psimd_s32 vm8 = vi8 > vmax;
182 vmax = psimd_blend_f32(vm8, vi8, vmax);
183 vidx = psimd_blend_u32(vm8, psimd_splat_u32(8), vidx);
184
185 psimd_f32 vout = psimd_max_f32(psimd_min_f32(vmax, voutput_max), voutput_min);
186
187 if (c & 2) {
188 psimd_store2_f32(output, vout);
189 psimd_store2_u32(index, vidx);
190 vout = psimd_concat_hi_f32(vout, vout);
191 vidx = psimd_concat_hi_u32(vidx, vidx);
192 output += 2;
193 index += 2;
194 }
195 if (c & 1) {
196 psimd_store1_f32(output, vout);
197 psimd_store1_u32(index, vidx);
198 output += 1;
199 index += 1;
200 }
201 }
202 input = (const float**) ((uintptr_t) input + input_increment);
203 output = (float*) ((uintptr_t) output + output_increment);
204 } while (--output_pixels != 0);
205 }
206