1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vmulcaddc/wasmsimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <wasm_simd128.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/vmulcaddc.h>
16
17
xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmrelaxedsimd_2x(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,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vmulcaddc_minmax_ukernel_c8__wasmrelaxedsimd_2x(
19 size_t rows,
20 size_t channels,
21 const float*restrict input,
22 size_t input_stride,
23 const float*restrict weights,
24 float*restrict output,
25 size_t output_stride,
26 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28 assert(rows != 0);
29 assert(channels != 0);
30 assert(channels % sizeof(float) == 0);
31
32 const float* i0 = input;
33 float* o0 = output;
34 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
35 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
36
37 const size_t input_increment = input_stride * 2 - channels;
38 const size_t output_increment = output_stride * 2 - channels;
39
40 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
41 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
42 do {
43 if XNN_UNPREDICTABLE(rows < 2) {
44 i1 = i0;
45 o1 = o0;
46 }
47
48 const float* w = weights;
49 size_t c = channels;
50 for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
51 const v128_t vscale0123 = wasm_v128_load(w);
52 const v128_t vscale4567 = wasm_v128_load(w + 4);
53
54 v128_t vacc0x0123 = wasm_v128_load(i0);
55 v128_t vacc0x4567 = wasm_v128_load(i0 + 4);
56 i0 += 8;
57 v128_t vacc1x0123 = wasm_v128_load(i1);
58 v128_t vacc1x4567 = wasm_v128_load(i1 + 4);
59 i1 += 8;
60
61 const v128_t vbias0123 = wasm_v128_load(w + 8);
62 const v128_t vbias4567 = wasm_v128_load(w + 12);
63
64 vacc0x0123 = wasm_f32x4_add(vbias0123, wasm_f32x4_mul(vscale0123, vacc0x0123));
65 vacc0x4567 = wasm_f32x4_add(vbias4567, wasm_f32x4_mul(vscale4567, vacc0x4567));
66 vacc1x0123 = wasm_f32x4_add(vbias0123, wasm_f32x4_mul(vscale0123, vacc1x0123));
67 vacc1x4567 = wasm_f32x4_add(vbias4567, wasm_f32x4_mul(vscale4567, vacc1x4567));
68
69 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
70 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
71 vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
72 vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
73
74 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
75 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
76 vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
77 vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
78
79 wasm_v128_store(o0, vacc0x0123);
80 wasm_v128_store(o0 + 4, vacc0x4567);
81 o0 += 8;
82 wasm_v128_store(o1, vacc1x0123);
83 wasm_v128_store(o1 + 4, vacc1x4567);
84 o1 += 8;
85
86 w += 16;
87 }
88 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
89 const v128_t vscale = wasm_v128_load(w);
90
91 v128_t vacc0 = wasm_v128_load(i0);
92 i0 += 4;
93 v128_t vacc1 = wasm_v128_load(i1);
94 i1 += 4;
95
96 const v128_t vbias = wasm_v128_load(w + 8);
97
98 vacc0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vscale, vacc0));
99 vacc1 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vscale, vacc1));
100
101 vacc0 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0);
102 vacc1 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1);
103
104 vacc0 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0);
105 vacc1 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1);
106
107 wasm_v128_store(o0, vacc0);
108 o0 += 4;
109 wasm_v128_store(o1, vacc1);
110 o1 += 4;
111
112 w += 4;
113 }
114 if XNN_UNLIKELY(c != 0) {
115 const v128_t vscale = wasm_v128_load(w);
116
117 v128_t vacc0 = wasm_v128_load(i0);
118 i0 = (const float*) ((uintptr_t) i0 + c);
119 v128_t vacc1 = wasm_v128_load(i1);
120 i1 = (const float*) ((uintptr_t) i1 + c);
121
122 const v128_t vbias = wasm_v128_load(w + 8);
123
124 vacc0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vscale, vacc0));
125 vacc1 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vscale, vacc1));
126
127 vacc0 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0);
128 vacc1 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1);
129
130 vacc0 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0);
131 vacc1 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1);
132
133 if (c & (2 * sizeof(float))) {
134 *((double*) o0) = wasm_f64x2_extract_lane(vacc0, 0);
135 *((double*) o1) = wasm_f64x2_extract_lane(vacc1, 0);
136
137 vacc0 = wasm_v32x4_shuffle(vacc0, vacc0, 2, 3, 2, 3);
138 vacc1 = wasm_v32x4_shuffle(vacc1, vacc1, 2, 3, 2, 3);
139
140 o0 += 2;
141 o1 += 2;
142 }
143 if (c & (1 * sizeof(float))) {
144 *o0++ = wasm_f32x4_extract_lane(vacc0, 0);
145 *o1++ = wasm_f32x4_extract_lane(vacc1, 0);
146 }
147 }
148 i0 = (const float*) ((uintptr_t) i0 + input_increment);
149 o0 = (float*) ((uintptr_t) o0 + output_increment);
150 i1 = (const float*) ((uintptr_t) i1 + input_increment);
151 o1 = (float*) ((uintptr_t) o1 + output_increment);
152 rows = doz(rows, 2);
153 } while (rows != 0);
154 }
155