• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_c1__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_c1__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     do {
48       const float vscale = w[0];
49 
50       float vacc0 = *i0++;
51       float vacc1 = *i1++;
52 
53       const float vbias = w[1];
54 
55       vacc0 = vacc0 * vscale + vbias;
56       vacc1 = vacc1 * vscale + vbias;
57 
58       vacc0 = __builtin_wasm_max_f32(vacc0, vmin);
59       vacc1 = __builtin_wasm_max_f32(vacc1, vmin);
60 
61       vacc0 = __builtin_wasm_min_f32(vacc0, vmax);
62       vacc1 = __builtin_wasm_min_f32(vacc1, vmax);
63 
64       *o0++ = vacc0;
65       *o1++ = vacc1;
66 
67       w += 2;
68       c -= sizeof(float);
69     } while (c != 0);
70     i0 = (const float*) ((uintptr_t) i0 + input_increment);
71     o0 = (float*) ((uintptr_t) o0 + output_increment);
72     i1 = (const float*) ((uintptr_t) i1 + input_increment);
73     o1 = (float*) ((uintptr_t) o1 + output_increment);
74     if XNN_UNPREDICTABLE(rows < 4) {
75       i1 = i0;
76       o1 = o0;
77     }
78     rows = doz(rows, 2);
79   } while (rows != 0);
80 }
81