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_2x1__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_2x1__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 - 2 * sizeof(float);
34 while (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 size_t n = nc;
39 while (n >= 1) {
40 uint32_t nnz = *nnzmap++;
41 float vacc0x0 = *w++;
42 float vacc1x0 = vacc0x0;
43 if XNN_LIKELY(nnz != 0) {
44 do {
45 const intptr_t diff = *dmap++;
46 const float vi0 = input[0];
47 const float vi1 = input[1];
48 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
49 const float vw0 = *w++;
50 vacc0x0 += vi0 * vw0;
51 vacc1x0 += vi1 * vw0;
52 } while (--nnz != 0);
53 }
54 float vout0x0 = math_min_f32(vacc0x0, vmax);
55 float vout1x0 = math_min_f32(vacc1x0, vmax);
56 vout0x0 = math_max_f32(vout0x0, vmin);
57 vout1x0 = math_max_f32(vout1x0, vmin);
58 output[0] = vout0x0;
59 output[1] = vout1x0;
60 output[0] = vout0x0;
61 output[1] = vout1x0;
62 output = (float*restrict) ((uintptr_t) output + output_stride);
63 n -= 1;
64 }
65 if XNN_UNLIKELY(n != 0) {
66 do {
67 uint32_t nnz = *nnzmap++;
68 float vacc0 = *w++;
69 float vacc1 = vacc0;
70 if XNN_LIKELY(nnz != 0) {
71 do {
72 const intptr_t diff = *dmap++;
73 const float vi0 = input[0];
74 const float vi1 = input[1];
75 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
76 const float vw = *w++;
77 vacc0 += vi0 * vw;
78 vacc1 += vi1 * vw;
79 } while (--nnz != 0);
80 }
81 float vout0 = math_min_f32(vacc0, vmax);
82 float vout1 = math_min_f32(vacc1, vmax);
83 vout0 = math_max_f32(vout0, vmin);
84 vout1 = math_max_f32(vout1, vmin);
85 output[0] = vout0;
86 output[1] = vout1;
87 output = (float*restrict) ((uintptr_t) output + output_stride);
88 n -= 1;
89 } while (n != 0);
90 }
91 output = (float*restrict) ((uintptr_t) output - output_decrement);
92 input += 2;
93 mc -= 2 * sizeof(float);
94 }
95 if XNN_UNLIKELY(mc != 0) {
96 output_decrement += 1 * sizeof(float);
97 if (mc & (1 * sizeof(float))) {
98 const float*restrict w = weights;
99 const int32_t* dmap = widx_dmap;
100 const uint32_t* nnzmap = nidx_nnzmap;
101 size_t n = nc;
102 while (n >= 1) {
103 uint32_t nnz = *nnzmap++;
104 float vacc0x0 = *w++;
105 if XNN_LIKELY(nnz != 0) {
106 do {
107 const intptr_t diff = *dmap++;
108 const float vi0 = input[0];
109 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
110 const float vw0 = *w++;
111 vacc0x0 += vi0 * vw0;
112 } while (--nnz != 0);
113 }
114 float vout0x0 = math_min_f32(vacc0x0, vmax);
115 vout0x0 = math_max_f32(vout0x0, vmin);
116 output[0] = vout0x0;
117 output = (float*restrict) ((uintptr_t) output + output_stride);
118 n -= 1;
119 }
120 if XNN_UNLIKELY(n != 0) {
121 do {
122 uint32_t nnz = *nnzmap++;
123 float vacc0 = *w++;
124 if XNN_LIKELY(nnz != 0) {
125 do {
126 const intptr_t diff = *dmap++;
127 const float vi0 = input[0];
128 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
129 const float vw = *w++;
130 vacc0 += vi0 * vw;
131 } while (--nnz != 0);
132 }
133 float vout0 = math_min_f32(vacc0, vmax);
134 vout0 = math_max_f32(vout0, vmin);
135 output[0] = vout0;
136 output = (float*restrict) ((uintptr_t) output + output_stride);
137 n -= 1;
138 } while (n != 0);
139 }
140 output = (float*restrict) ((uintptr_t) output - output_decrement);
141 input += 1;
142 }
143 }
144 }
145