• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-spmm/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/spmm.h>
14 
15 
xnn_f32_spmm_minmax_ukernel_1x1__scalar(size_t mc,size_t nc,const float * restrict input,const float * restrict weights,const int32_t * restrict widx_dmap,const uint32_t * restrict nidx_nnzmap,float * restrict output,size_t output_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_spmm_minmax_ukernel_1x1__scalar(
17     size_t mc,
18     size_t nc,
19     const float*restrict input,
20     const float*restrict weights,
21     const int32_t*restrict widx_dmap,
22     const uint32_t*restrict nidx_nnzmap,
23     float*restrict output,
24     size_t output_stride,
25     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
26 {
27   assert(mc != 0);
28   assert(mc % sizeof(float) == 0);
29   assert(nc != 0);
30 
31   const float vmin = params->scalar.min;
32   const float vmax = params->scalar.max;
33   size_t output_decrement = output_stride * nc - 1 * sizeof(float);
34   while (mc >= 1 * sizeof(float)) {
35     const float*restrict w = weights;
36     const int32_t* dmap = widx_dmap;
37     const uint32_t* nnzmap = nidx_nnzmap;
38     size_t n = nc;
39     while (n >= 1) {
40       uint32_t nnz = *nnzmap++;
41       float vacc0x0 = *w++;
42       if XNN_LIKELY(nnz != 0) {
43         do {
44           const intptr_t diff = *dmap++;
45           const float vi0 = input[0];
46           input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
47           const float vw0 = *w++;
48           vacc0x0 += vi0 * vw0;
49         } while (--nnz != 0);
50       }
51       float vout0x0 = math_min_f32(vacc0x0, vmax);
52       vout0x0 = math_max_f32(vout0x0, vmin);
53       output[0] = vout0x0;
54       output[0] = vout0x0;
55       output = (float*restrict) ((uintptr_t) output + output_stride);
56       n -= 1;
57     }
58     if XNN_UNLIKELY(n != 0) {
59       do {
60         uint32_t nnz = *nnzmap++;
61         float vacc0 = *w++;
62         if XNN_LIKELY(nnz != 0) {
63           do {
64             const intptr_t diff = *dmap++;
65             const float vi0 = input[0];
66             input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
67             const float vw = *w++;
68             vacc0 += vi0 * vw;
69           } while (--nnz != 0);
70         }
71         float vout0 = math_min_f32(vacc0, vmax);
72         vout0 = math_max_f32(vout0, vmin);
73         output[0] = vout0;
74         output = (float*restrict) ((uintptr_t) output + output_stride);
75         n -= 1;
76       } while (n != 0);
77     }
78     output = (float*restrict) ((uintptr_t) output - output_decrement);
79     input += 1;
80     mc -= 1 * sizeof(float);
81   }
82   if XNN_UNLIKELY(mc != 0) {
83   }
84 }
85