1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vmulcaddc/scalar.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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 <xnnpack/math.h>
13 #include <xnnpack/vmulcaddc.h>
14
15
xnn_f32_vmulcaddc_minmax_ukernel_c2__wasm_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)])16 void xnn_f32_vmulcaddc_minmax_ukernel_c2__wasm_2x(
17 size_t rows,
18 size_t channels,
19 const float*restrict input,
20 size_t input_stride,
21 const float*restrict weights,
22 float*restrict output,
23 size_t output_stride,
24 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(rows != 0);
27 assert(channels != 0);
28 assert(channels % sizeof(float) == 0);
29
30 const size_t input_increment = input_stride * 2 - channels;
31 const size_t output_increment = output_stride * 2 - channels;
32
33 const float* i0 = input;
34 float* o0 = output;
35 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
36 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
37 if XNN_UNPREDICTABLE(rows < 2) {
38 i1 = i0;
39 o1 = o0;
40 }
41
42 const float vmin = params->scalar.min;
43 const float vmax = params->scalar.max;
44 do {
45 const float* w = weights;
46 size_t c = channels;
47 for (; c >= 2 * sizeof(float); c -= 2 * sizeof(float)) {
48 const float vscale0 = w[0];
49 const float vscale1 = w[1];
50
51 float vacc0x0 = i0[0];
52 float vacc0x1 = i0[1];
53 i0 += 2;
54 float vacc1x0 = i1[0];
55 float vacc1x1 = i1[1];
56 i1 += 2;
57
58 const float vbias0 = w[2];
59 const float vbias1 = w[3];
60
61 vacc0x0 = vacc0x0 * vscale0 + vbias0;
62 vacc0x1 = vacc0x1 * vscale1 + vbias1;
63 vacc1x0 = vacc1x0 * vscale0 + vbias0;
64 vacc1x1 = vacc1x1 * vscale1 + vbias1;
65
66 vacc0x0 = __builtin_wasm_max_f32(vacc0x0, vmin);
67 vacc0x1 = __builtin_wasm_max_f32(vacc0x1, vmin);
68 vacc1x0 = __builtin_wasm_max_f32(vacc1x0, vmin);
69 vacc1x1 = __builtin_wasm_max_f32(vacc1x1, vmin);
70
71 vacc0x0 = __builtin_wasm_min_f32(vacc0x0, vmax);
72 vacc0x1 = __builtin_wasm_min_f32(vacc0x1, vmax);
73 vacc1x0 = __builtin_wasm_min_f32(vacc1x0, vmax);
74 vacc1x1 = __builtin_wasm_min_f32(vacc1x1, vmax);
75
76 o0[0] = vacc0x0;
77 o0[1] = vacc0x1;
78 o0 += 2;
79 o1[0] = vacc1x0;
80 o1[1] = vacc1x1;
81 o1 += 2;
82
83 w += 4;
84 }
85 if XNN_UNLIKELY(c != 0) {
86 do {
87 const float vscale = *w++;
88
89 float vacc0 = *i0++;
90 float vacc1 = *i1++;
91
92 const float vbias = w[1];
93
94 vacc0 = vacc0 * vscale + vbias;
95 vacc1 = vacc1 * vscale + vbias;
96
97 vacc0 = __builtin_wasm_max_f32(vacc0, vmin);
98 vacc1 = __builtin_wasm_max_f32(vacc1, vmin);
99
100 vacc0 = __builtin_wasm_min_f32(vacc0, vmax);
101 vacc1 = __builtin_wasm_min_f32(vacc1, vmax);
102
103 *o0++ = vacc0;
104 *o1++ = vacc1;
105
106 c -= sizeof(float);
107 } while (c != 0);
108 }
109 i0 = (const float*) ((uintptr_t) i0 + input_increment);
110 o0 = (float*) ((uintptr_t) o0 + output_increment);
111 i1 = (const float*) ((uintptr_t) i1 + input_increment);
112 o1 = (float*) ((uintptr_t) o1 + output_increment);
113 if XNN_UNPREDICTABLE(rows < 4) {
114 i1 = i0;
115 o1 = o0;
116 }
117 rows = doz(rows, 2);
118 } while (rows != 0);
119 }
120