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_4x1__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_4x1__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 - 4 * sizeof(float);
34 while (mc >= 4 * 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 float vacc2x0 = vacc0x0;
44 float vacc3x0 = vacc0x0;
45 if XNN_LIKELY(nnz != 0) {
46 do {
47 const intptr_t diff = *dmap++;
48 const float vi0 = input[0];
49 const float vi1 = input[1];
50 const float vi2 = input[2];
51 const float vi3 = input[3];
52 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
53 const float vw0 = *w++;
54 vacc0x0 += vi0 * vw0;
55 vacc1x0 += vi1 * vw0;
56 vacc2x0 += vi2 * vw0;
57 vacc3x0 += vi3 * vw0;
58 } while (--nnz != 0);
59 }
60 float vout0x0 = math_min_f32(vacc0x0, vmax);
61 float vout1x0 = math_min_f32(vacc1x0, vmax);
62 float vout2x0 = math_min_f32(vacc2x0, vmax);
63 float vout3x0 = math_min_f32(vacc3x0, vmax);
64 vout0x0 = math_max_f32(vout0x0, vmin);
65 vout1x0 = math_max_f32(vout1x0, vmin);
66 vout2x0 = math_max_f32(vout2x0, vmin);
67 vout3x0 = math_max_f32(vout3x0, vmin);
68 output[0] = vout0x0;
69 output[1] = vout1x0;
70 output[2] = vout2x0;
71 output[3] = vout3x0;
72 output[0] = vout0x0;
73 output[1] = vout1x0;
74 output[2] = vout2x0;
75 output[3] = vout3x0;
76 output = (float*restrict) ((uintptr_t) output + output_stride);
77 n -= 1;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 do {
81 uint32_t nnz = *nnzmap++;
82 float vacc0 = *w++;
83 float vacc1 = vacc0;
84 float vacc2 = vacc0;
85 float vacc3 = vacc0;
86 if XNN_LIKELY(nnz != 0) {
87 do {
88 const intptr_t diff = *dmap++;
89 const float vi0 = input[0];
90 const float vi1 = input[1];
91 const float vi2 = input[2];
92 const float vi3 = input[3];
93 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
94 const float vw = *w++;
95 vacc0 += vi0 * vw;
96 vacc1 += vi1 * vw;
97 vacc2 += vi2 * vw;
98 vacc3 += vi3 * vw;
99 } while (--nnz != 0);
100 }
101 float vout0 = math_min_f32(vacc0, vmax);
102 float vout1 = math_min_f32(vacc1, vmax);
103 float vout2 = math_min_f32(vacc2, vmax);
104 float vout3 = math_min_f32(vacc3, vmax);
105 vout0 = math_max_f32(vout0, vmin);
106 vout1 = math_max_f32(vout1, vmin);
107 vout2 = math_max_f32(vout2, vmin);
108 vout3 = math_max_f32(vout3, vmin);
109 output[0] = vout0;
110 output[1] = vout1;
111 output[2] = vout2;
112 output[3] = vout3;
113 output = (float*restrict) ((uintptr_t) output + output_stride);
114 n -= 1;
115 } while (n != 0);
116 }
117 output = (float*restrict) ((uintptr_t) output - output_decrement);
118 input += 4;
119 mc -= 4 * sizeof(float);
120 }
121 if XNN_UNLIKELY(mc != 0) {
122 output_decrement += 2 * sizeof(float);
123 if (mc & (2 * sizeof(float))) {
124 const float*restrict w = weights;
125 const int32_t* dmap = widx_dmap;
126 const uint32_t* nnzmap = nidx_nnzmap;
127 size_t n = nc;
128 while (n >= 1) {
129 uint32_t nnz = *nnzmap++;
130 float vacc0x0 = *w++;
131 float vacc1x0 = vacc0x0;
132 if XNN_LIKELY(nnz != 0) {
133 do {
134 const intptr_t diff = *dmap++;
135 const float vi0 = input[0];
136 const float vi1 = input[1];
137 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
138 const float vw0 = *w++;
139 vacc0x0 += vi0 * vw0;
140 vacc1x0 += vi1 * vw0;
141 } while (--nnz != 0);
142 }
143 float vout0x0 = math_min_f32(vacc0x0, vmax);
144 float vout1x0 = math_min_f32(vacc1x0, vmax);
145 vout0x0 = math_max_f32(vout0x0, vmin);
146 vout1x0 = math_max_f32(vout1x0, vmin);
147 output[0] = vout0x0;
148 output[1] = vout1x0;
149 output = (float*restrict) ((uintptr_t) output + output_stride);
150 n -= 1;
151 }
152 if XNN_UNLIKELY(n != 0) {
153 do {
154 uint32_t nnz = *nnzmap++;
155 float vacc0 = *w++;
156 float vacc1 = vacc0;
157 if XNN_LIKELY(nnz != 0) {
158 do {
159 const intptr_t diff = *dmap++;
160 const float vi0 = input[0];
161 const float vi1 = input[1];
162 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
163 const float vw = *w++;
164 vacc0 += vi0 * vw;
165 vacc1 += vi1 * vw;
166 } while (--nnz != 0);
167 }
168 float vout0 = math_min_f32(vacc0, vmax);
169 float vout1 = math_min_f32(vacc1, vmax);
170 vout0 = math_max_f32(vout0, vmin);
171 vout1 = math_max_f32(vout1, vmin);
172 output[0] = vout0;
173 output[1] = vout1;
174 output = (float*restrict) ((uintptr_t) output + output_stride);
175 n -= 1;
176 } while (n != 0);
177 }
178 output = (float*restrict) ((uintptr_t) output - output_decrement);
179 input += 2;
180 }
181 output_decrement += 1 * sizeof(float);
182 if (mc & (1 * sizeof(float))) {
183 const float*restrict w = weights;
184 const int32_t* dmap = widx_dmap;
185 const uint32_t* nnzmap = nidx_nnzmap;
186 size_t n = nc;
187 while (n >= 1) {
188 uint32_t nnz = *nnzmap++;
189 float vacc0x0 = *w++;
190 if XNN_LIKELY(nnz != 0) {
191 do {
192 const intptr_t diff = *dmap++;
193 const float vi0 = input[0];
194 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
195 const float vw0 = *w++;
196 vacc0x0 += vi0 * vw0;
197 } while (--nnz != 0);
198 }
199 float vout0x0 = math_min_f32(vacc0x0, vmax);
200 vout0x0 = math_max_f32(vout0x0, vmin);
201 output[0] = vout0x0;
202 output = (float*restrict) ((uintptr_t) output + output_stride);
203 n -= 1;
204 }
205 if XNN_UNLIKELY(n != 0) {
206 do {
207 uint32_t nnz = *nnzmap++;
208 float vacc0 = *w++;
209 if XNN_LIKELY(nnz != 0) {
210 do {
211 const intptr_t diff = *dmap++;
212 const float vi0 = input[0];
213 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
214 const float vw = *w++;
215 vacc0 += vi0 * vw;
216 } while (--nnz != 0);
217 }
218 float vout0 = math_min_f32(vacc0, vmax);
219 vout0 = math_max_f32(vout0, vmin);
220 output[0] = vout0;
221 output = (float*restrict) ((uintptr_t) output + output_stride);
222 n -= 1;
223 } while (n != 0);
224 }
225 output = (float*restrict) ((uintptr_t) output - output_decrement);
226 input += 1;
227 }
228 }
229 }
230