1 // Auto-generated file. Do not edit!
2 // Template: src/f32-spmm/scalar-pipelined.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_2x1__scalar_pipelined(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_2x1__scalar_pipelined(
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 - 2 * sizeof(float);
34 while XNN_LIKELY(mc >= 2 * sizeof(float)) {
35 const float*restrict w = weights;
36 const int32_t* dmap = widx_dmap;
37 const uint32_t* nnzmap = nidx_nnzmap;
38 float vw = *w++;
39 intptr_t diff = *dmap++;
40 float vi0 = input[0];
41 float vi1 = input[1];
42 size_t n = nc;
43 do {
44 uint32_t nnz = *nnzmap++;
45 float vacc0 = vw;
46 float vacc1 = vw;
47 vw = *w++;
48 if XNN_LIKELY(nnz != 0) {
49 do {
50 vacc0 += vi0 * vw;
51 vacc1 += vi1 * vw;
52 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
53
54 diff = *dmap++;
55 vw = *w++;
56 vi0 = input[0];
57 vi1 = input[1];
58 } while (--nnz != 0);
59 }
60 float vout0 = math_min_f32(vacc0, vmax);
61 float vout1 = math_min_f32(vacc1, vmax);
62 vout0 = math_max_f32(vout0, vmin);
63 vout1 = math_max_f32(vout1, vmin);
64 output[0] = vout0;
65 output[1] = vout1;
66 output = (float*restrict) ((uintptr_t) output + output_stride);
67 } while (--n != 0);
68 output = (float*restrict) ((uintptr_t) output - output_decrement);
69 input += 2;
70 mc -= 2 * sizeof(float);
71 }
72 if XNN_UNLIKELY(mc != 0) {
73 output_decrement += 1 * sizeof(float);
74 if (mc & (1 * sizeof(float))) {
75 const float*restrict w = weights;
76 const int32_t* dmap = widx_dmap;
77 const uint32_t* nnzmap = nidx_nnzmap;
78 float vw = *w++;
79 intptr_t diff = *dmap++;
80 float vi0 = input[0];
81 size_t n = nc;
82 do {
83 uint32_t nnz = *nnzmap++;
84 float vacc0 = vw;
85 vw = *w++;
86 if XNN_LIKELY(nnz != 0) {
87 do {
88 vacc0 += vi0 * vw;
89 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
90
91 diff = *dmap++;
92 vw = *w++;
93 vi0 = input[0];
94 } while (--nnz != 0);
95 }
96 float vout0 = math_min_f32(vacc0, vmax);
97 vout0 = math_max_f32(vout0, vmin);
98 output[0] = vout0;
99 output = (float*restrict) ((uintptr_t) output + output_stride);
100 } while (--n != 0);
101 output = (float*restrict) ((uintptr_t) output - output_decrement);
102 input += 1;
103 }
104 }
105 }
106