1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/f32-prelu/wasmsimd-minmax.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker #include <wasm_simd128.h>
13*4bdc9457SAndroid Build Coastguard Worker
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/prelu.h>
16*4bdc9457SAndroid Build Coastguard Worker
17*4bdc9457SAndroid Build Coastguard Worker
xnn_f32_prelu_ukernel__wasmsimd_minmax_2x8(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride)18*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_prelu_ukernel__wasmsimd_minmax_2x8(
19*4bdc9457SAndroid Build Coastguard Worker size_t rows,
20*4bdc9457SAndroid Build Coastguard Worker size_t channels,
21*4bdc9457SAndroid Build Coastguard Worker const float*restrict input,
22*4bdc9457SAndroid Build Coastguard Worker size_t input_stride,
23*4bdc9457SAndroid Build Coastguard Worker const float*restrict weights,
24*4bdc9457SAndroid Build Coastguard Worker float*restrict output,
25*4bdc9457SAndroid Build Coastguard Worker size_t output_stride) XNN_OOB_READS
26*4bdc9457SAndroid Build Coastguard Worker {
27*4bdc9457SAndroid Build Coastguard Worker assert(rows != 0);
28*4bdc9457SAndroid Build Coastguard Worker assert(channels != 0);
29*4bdc9457SAndroid Build Coastguard Worker assert(channels % sizeof(float) == 0);
30*4bdc9457SAndroid Build Coastguard Worker
31*4bdc9457SAndroid Build Coastguard Worker const float* i0 = input;
32*4bdc9457SAndroid Build Coastguard Worker float* o0 = output;
33*4bdc9457SAndroid Build Coastguard Worker const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
34*4bdc9457SAndroid Build Coastguard Worker float* o1 = (float*) ((uintptr_t) o0 + output_stride);
35*4bdc9457SAndroid Build Coastguard Worker
36*4bdc9457SAndroid Build Coastguard Worker const size_t input_increment = input_stride * 2 - channels;
37*4bdc9457SAndroid Build Coastguard Worker const size_t output_increment = output_stride * 2 - channels;
38*4bdc9457SAndroid Build Coastguard Worker
39*4bdc9457SAndroid Build Coastguard Worker const v128_t vzero = wasm_i32x4_const_splat(0);
40*4bdc9457SAndroid Build Coastguard Worker do {
41*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(rows < 2) {
42*4bdc9457SAndroid Build Coastguard Worker i1 = i0;
43*4bdc9457SAndroid Build Coastguard Worker o1 = o0;
44*4bdc9457SAndroid Build Coastguard Worker }
45*4bdc9457SAndroid Build Coastguard Worker
46*4bdc9457SAndroid Build Coastguard Worker const float* w = weights;
47*4bdc9457SAndroid Build Coastguard Worker size_t c = channels;
48*4bdc9457SAndroid Build Coastguard Worker for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
49*4bdc9457SAndroid Build Coastguard Worker const v128_t vw0123 = wasm_v128_load(w);
50*4bdc9457SAndroid Build Coastguard Worker const v128_t vw4567 = wasm_v128_load(w + 4);
51*4bdc9457SAndroid Build Coastguard Worker w += 8;
52*4bdc9457SAndroid Build Coastguard Worker
53*4bdc9457SAndroid Build Coastguard Worker v128_t vi0x0123 = wasm_v128_load(i0);
54*4bdc9457SAndroid Build Coastguard Worker v128_t vi0x4567 = wasm_v128_load(i0 + 4);
55*4bdc9457SAndroid Build Coastguard Worker i0 += 8;
56*4bdc9457SAndroid Build Coastguard Worker v128_t vi1x0123 = wasm_v128_load(i1);
57*4bdc9457SAndroid Build Coastguard Worker v128_t vi1x4567 = wasm_v128_load(i1 + 4);
58*4bdc9457SAndroid Build Coastguard Worker i1 += 8;
59*4bdc9457SAndroid Build Coastguard Worker
60*4bdc9457SAndroid Build Coastguard Worker v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
61*4bdc9457SAndroid Build Coastguard Worker vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
62*4bdc9457SAndroid Build Coastguard Worker v128_t vacc0x4567 = wasm_i32x4_max(vi0x4567, vzero);
63*4bdc9457SAndroid Build Coastguard Worker vi0x4567 = wasm_i32x4_min(vi0x4567, vzero);
64*4bdc9457SAndroid Build Coastguard Worker v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
65*4bdc9457SAndroid Build Coastguard Worker vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
66*4bdc9457SAndroid Build Coastguard Worker v128_t vacc1x4567 = wasm_i32x4_max(vi1x4567, vzero);
67*4bdc9457SAndroid Build Coastguard Worker vi1x4567 = wasm_i32x4_min(vi1x4567, vzero);
68*4bdc9457SAndroid Build Coastguard Worker
69*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
70*4bdc9457SAndroid Build Coastguard Worker vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(vi0x4567, vw4567));
71*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
72*4bdc9457SAndroid Build Coastguard Worker vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(vi1x4567, vw4567));
73*4bdc9457SAndroid Build Coastguard Worker
74*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(o0, vacc0x0123);
75*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(o0 + 4, vacc0x4567);
76*4bdc9457SAndroid Build Coastguard Worker o0 += 8;
77*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(o1, vacc1x0123);
78*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(o1 + 4, vacc1x4567);
79*4bdc9457SAndroid Build Coastguard Worker o1 += 8;
80*4bdc9457SAndroid Build Coastguard Worker }
81*4bdc9457SAndroid Build Coastguard Worker for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
82*4bdc9457SAndroid Build Coastguard Worker const v128_t vw0123 = wasm_v128_load(w);
83*4bdc9457SAndroid Build Coastguard Worker w += 4;
84*4bdc9457SAndroid Build Coastguard Worker
85*4bdc9457SAndroid Build Coastguard Worker v128_t vi0x0123 = wasm_v128_load(i0);
86*4bdc9457SAndroid Build Coastguard Worker i0 += 4;
87*4bdc9457SAndroid Build Coastguard Worker v128_t vi1x0123 = wasm_v128_load(i1);
88*4bdc9457SAndroid Build Coastguard Worker i1 += 4;
89*4bdc9457SAndroid Build Coastguard Worker
90*4bdc9457SAndroid Build Coastguard Worker v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
91*4bdc9457SAndroid Build Coastguard Worker vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
92*4bdc9457SAndroid Build Coastguard Worker v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
93*4bdc9457SAndroid Build Coastguard Worker vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
94*4bdc9457SAndroid Build Coastguard Worker
95*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
96*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
97*4bdc9457SAndroid Build Coastguard Worker
98*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(o0, vacc0x0123);
99*4bdc9457SAndroid Build Coastguard Worker o0 += 4;
100*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(o1, vacc1x0123);
101*4bdc9457SAndroid Build Coastguard Worker o1 += 4;
102*4bdc9457SAndroid Build Coastguard Worker }
103*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(c != 0) {
104*4bdc9457SAndroid Build Coastguard Worker const v128_t vw0123 = wasm_v128_load(w);
105*4bdc9457SAndroid Build Coastguard Worker w = (const float*) ((uintptr_t) w + c);
106*4bdc9457SAndroid Build Coastguard Worker
107*4bdc9457SAndroid Build Coastguard Worker v128_t vi0x0123 = wasm_v128_load(i0);
108*4bdc9457SAndroid Build Coastguard Worker i0 = (const float*) ((uintptr_t) i0 + c);
109*4bdc9457SAndroid Build Coastguard Worker v128_t vi1x0123 = wasm_v128_load(i1);
110*4bdc9457SAndroid Build Coastguard Worker i1 = (const float*) ((uintptr_t) i1 + c);
111*4bdc9457SAndroid Build Coastguard Worker
112*4bdc9457SAndroid Build Coastguard Worker v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
113*4bdc9457SAndroid Build Coastguard Worker vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
114*4bdc9457SAndroid Build Coastguard Worker v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
115*4bdc9457SAndroid Build Coastguard Worker vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
116*4bdc9457SAndroid Build Coastguard Worker
117*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
118*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
119*4bdc9457SAndroid Build Coastguard Worker
120*4bdc9457SAndroid Build Coastguard Worker if (c & (2 * sizeof(float))) {
121*4bdc9457SAndroid Build Coastguard Worker *((double*) o0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
122*4bdc9457SAndroid Build Coastguard Worker *((double*) o1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
123*4bdc9457SAndroid Build Coastguard Worker
124*4bdc9457SAndroid Build Coastguard Worker vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
125*4bdc9457SAndroid Build Coastguard Worker vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
126*4bdc9457SAndroid Build Coastguard Worker
127*4bdc9457SAndroid Build Coastguard Worker o0 += 2;
128*4bdc9457SAndroid Build Coastguard Worker o1 += 2;
129*4bdc9457SAndroid Build Coastguard Worker }
130*4bdc9457SAndroid Build Coastguard Worker if (c & (1 * sizeof(float))) {
131*4bdc9457SAndroid Build Coastguard Worker *o0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
132*4bdc9457SAndroid Build Coastguard Worker *o1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
133*4bdc9457SAndroid Build Coastguard Worker
134*4bdc9457SAndroid Build Coastguard Worker o0 += 1;
135*4bdc9457SAndroid Build Coastguard Worker o1 += 1;
136*4bdc9457SAndroid Build Coastguard Worker }
137*4bdc9457SAndroid Build Coastguard Worker }
138*4bdc9457SAndroid Build Coastguard Worker i0 = (const float*) ((uintptr_t) i0 + input_increment);
139*4bdc9457SAndroid Build Coastguard Worker o0 = (float*) ((uintptr_t) o0 + output_increment);
140*4bdc9457SAndroid Build Coastguard Worker i1 = (const float*) ((uintptr_t) i1 + input_increment);
141*4bdc9457SAndroid Build Coastguard Worker o1 = (float*) ((uintptr_t) o1 + output_increment);
142*4bdc9457SAndroid Build Coastguard Worker rows = doz(rows, 2);
143*4bdc9457SAndroid Build Coastguard Worker } while (rows != 0);
144*4bdc9457SAndroid Build Coastguard Worker }
145