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/gavgpool.h>
11
12
xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_x86_c4(size_t rows,size_t channels,const float * input,size_t input_stride,const float * zero,float * output,const union xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])13 void xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_x86_c4(
14 size_t rows,
15 size_t channels,
16 const float* input,
17 size_t input_stride,
18 const float* zero,
19 float* output,
20 const union xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
21 {
22 assert(rows != 0);
23 assert(rows <= 7);
24 assert(channels != 0);
25
26 const float* i0 = input;
27 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
28 if (rows < 2) {
29 i1 = zero;
30 }
31 const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
32 if (rows <= 2) {
33 i2 = zero;
34 }
35 const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
36 if (rows < 4) {
37 i3 = zero;
38 }
39 const float* i4 = (const float*) ((uintptr_t) i3 + input_stride);
40 if (rows <= 4) {
41 i4 = zero;
42 }
43 const float* i5 = (const float*) ((uintptr_t) i4 + input_stride);
44 if (rows < 6) {
45 i5 = zero;
46 }
47 const float* i6 = (const float*) ((uintptr_t) i5 + input_stride);
48 if (rows <= 6) {
49 i6 = zero;
50 }
51 const v128_t vscale = wasm_v128_load32_splat(¶ms->scalar.scale);
52 const v128_t vmin = wasm_v128_load32_splat(¶ms->scalar.min);
53 const v128_t vmax = wasm_v128_load32_splat(¶ms->scalar.max);
54
55 while (channels >= 4) {
56 const v128_t vi0 = wasm_v128_load(i0);
57 i0 += 4;
58 const v128_t vi1 = wasm_v128_load(i1);
59 i1 += 4;
60 const v128_t vi2 = wasm_v128_load(i2);
61 i2 += 4;
62 const v128_t vi3 = wasm_v128_load(i3);
63 i3 += 4;
64 const v128_t vi4 = wasm_v128_load(i4);
65 i4 += 4;
66 const v128_t vi5 = wasm_v128_load(i5);
67 i5 += 4;
68 const v128_t vi6 = wasm_v128_load(i6);
69 i6 += 4;
70
71 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1);
72 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3);
73 const v128_t vsum45 = wasm_f32x4_add(vi4, vi5);
74
75 const v128_t vsum016 = wasm_f32x4_add(vsum01, vi6);
76 const v128_t vsum2345 = wasm_f32x4_add(vsum23, vsum45);
77
78 const v128_t vsum = wasm_f32x4_add(vsum016, vsum2345);
79
80 v128_t vout = wasm_f32x4_mul(vsum, vscale);
81 vout = wasm_f32x4_pmax(vmin, vout);
82 vout = wasm_f32x4_pmin(vmax, vout);
83
84 wasm_v128_store(output, vout);
85 output += 4;
86
87 channels -= 4;
88 }
89 if (channels != 0) {
90 const v128_t vi0 = wasm_v128_load(i0);
91 const v128_t vi1 = wasm_v128_load(i1);
92 const v128_t vi2 = wasm_v128_load(i2);
93 const v128_t vi3 = wasm_v128_load(i3);
94 const v128_t vi4 = wasm_v128_load(i4);
95 const v128_t vi5 = wasm_v128_load(i5);
96 const v128_t vi6 = wasm_v128_load(i6);
97
98 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1);
99 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3);
100 const v128_t vsum45 = wasm_f32x4_add(vi4, vi5);
101
102 const v128_t vsum016 = wasm_f32x4_add(vsum01, vi6);
103 const v128_t vsum2345 = wasm_f32x4_add(vsum23, vsum45);
104
105 const v128_t vsum = wasm_f32x4_add(vsum016, vsum2345);
106
107 v128_t vout = wasm_f32x4_mul(vsum, vscale);
108 vout = wasm_f32x4_pmax(vmin, vout);
109 vout = wasm_f32x4_pmin(vmax, vout);
110
111 if (channels & 2) {
112 *((double*) output) = wasm_f64x2_extract_lane(vout, 0);
113 output += 2;
114 vout = wasm_v32x4_shuffle(vout, vout, 2, 3, 2, 3);
115 }
116 if (channels & 1) {
117 *output++ = wasm_f32x4_extract_lane(vout, 0);
118 }
119 }
120 }
121