-
Notifications
You must be signed in to change notification settings - Fork 2
Expand file tree
/
Copy pathproject.h
More file actions
329 lines (220 loc) · 12.2 KB
/
project.h
File metadata and controls
329 lines (220 loc) · 12.2 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
#include <vector>
#include <adf.h>
#include "kernels.h"
#include "include.h"
using namespace adf;
class simpleGraph : public adf::graph {
private:
kernel mat_mul_k[mult_Y * mult_X * mult_Z];
kernel add_k[(mult_Y - 1) * mult_X * mult_Z];
public:
input_plio A[mult_X * mult_Y];
input_plio B[mult_Y * mult_Z];
output_plio C[mult_X * mult_Z];
simpleGraph(){
// input and output PLIOs creation below
// A: 0 1 2 ...
// x x+1 x+2 ...
for (int i = 0; i < mult_X * mult_Y; i++){
A[i] = input_plio::create(plio_128_bits, "data/matA" + std::to_string(i) + ".txt");
}
// B: 0 y
// 1 y+1
// 2 y+2
// ... ...
for (int i = 0; i < mult_Y * mult_Z; i++){
B[i] = input_plio::create(plio_128_bits, "data/matB" + std::to_string(i) + ".txt");
}
// C: 0 1 2 ...
// x x+1 x+2 ...
for (int i = 0; i < mult_X * mult_Z; i++){
C[i] = output_plio::create(plio_128_bits, "data/matC" + std::to_string(i) + ".txt");
}
// kernels creation
for (int i = 0; i < mult_Y * mult_X * mult_Z; i++){
mat_mul_k[i] = kernel::create(opt_blocked_matrix_mult);
}
for (int i = 0; i < (mult_Y - 1) * mult_X * mult_Z; i++){
add_k[i] = kernel::create(vectorized_add);
}
// automated graph generation
for (int i = 0; i < mult_X; i++){
for (int j = 0; j < mult_Z; j++){
for (int k = 0; k < mult_Y; k++){
// inputs (PL) --> mat mul kernels connection
connect< window<single_M*single_K*1> > (A[i*mult_Y + k].out[0], mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + k].in[0]);
connect< window<single_K*single_N*1> > (B[j*mult_Y + k].out[0], mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + k].in[1]);
// Place buffers in different banks to prevent memory stalls (see UG1076 for more details)
not_equal(location<buffer>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + k].in[0]), location<buffer>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + k].in[1]));
}
// adder tree automatic generation
// this vector holds the nodes to be added in the tree that haven't pair with others yet.
// Maximum number of entries is 2 (any 2 should pair together when possible)
vector<kernel> nodes_to_add = {};
int previous_k = mult_Y;
int k = mult_Y/2;
int iter = 0;
int adder_counter = 0;
int prev_layer_counter = 0;
while (k > 0){
if (iter == 0){ // first level of the tree
for (int l = 0; l < k; l++){
connect< window<single_M*single_N*4> > (mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 2*l].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[0]);
connect< window<single_M*single_N*4> > (mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 2*l + 1].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[1]);
// Place buffers in different banks to prevent memory stalls (see UG1076 for more details)
not_equal(location<buffer>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[0]), location<buffer>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[1]));
adder_counter++;
}
prev_layer_counter = adder_counter;
// if one node is not yet part of the adder tree,
// add it to vector to pair it with another similar node
if (2*k < previous_k){
nodes_to_add.push_back(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + previous_k - 1]);
}
}
else { // next level of the tree
int current_counter = 0;
for (int l = 0; l < k; l++){
connect< window<single_M*single_N*4> > (add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter - prev_layer_counter + l].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[0]);
connect< window<single_M*single_N*4> > (add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter - prev_layer_counter + l + 1].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[1]);
adder_counter++;
current_counter++;
}
// if one node is not yet part of the adder tree,
// add it to vector to pair it with another similar node
if (2*k < previous_k){
nodes_to_add.push_back(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter - current_counter - 1]);
}
// check wheather 2 nodeds have been added to vector
// if there are 2, put an adder and remove from vector
if (nodes_to_add.size() == 2){
connect< window<single_M*single_N*4> > (nodes_to_add[0].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[0]);
connect< window<single_M*single_N*4> > (nodes_to_add[1].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[1]);
// remove nodes
nodes_to_add.pop_back();
nodes_to_add.pop_back();
adder_counter++;
current_counter++;
}
prev_layer_counter = current_counter;
}
// next iteration
iter ++;
previous_k = prev_layer_counter;
k = prev_layer_counter/2;
}
// if counter of previous layer is > 1
// this means that it should be 2, and the adder came from vector
if (prev_layer_counter == 2){
connect< window<single_M*single_N*4> > (add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter - prev_layer_counter].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[0]);
connect< window<single_M*single_N*4> > (add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter - prev_layer_counter + 1].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[1]);
adder_counter++;
}
// there is one remaining node at the vector
// put an adder to it
else if (nodes_to_add.size() > 0){
connect< window<single_M*single_N*4> > (add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter - prev_layer_counter].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[0]);
connect< window<single_M*single_N*4> > (nodes_to_add[0].out[0], add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter].in[1]);
nodes_to_add.pop_back();
adder_counter++;
}
// adder --> output (PL) connection
connect< window<single_M*single_N*4> > (add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + adder_counter - 1].out[0], C[i*mult_Z + j].in[0]);
}
}
// direct the source file of kernels
for (int i = 0; i < mult_Y * mult_X * mult_Z; i++){
source(mat_mul_k[i]) = "kernels/kernels.cc";
}
for (int i = 0; i < (mult_Y - 1) * mult_X * mult_Z; i++){
source(add_k[i]) = "kernels/kernels.cc";
}
int top_x_offset = 0;
// runtime ratio and place constraints for MatMul kernels
for (int i = 0; i < mult_X; i++){
for (int j = 0; j < mult_Z; j++){
// for (int k = 0; k < mult_Y; k++){
runtime<ratio>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 0]) = 1.0;
runtime<ratio>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 1]) = 1.0;
runtime<ratio>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 2]) = 1.0;
runtime<ratio>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 3]) = 1.0;
if (((i*mult_Z + j)%3 == 0) && (((i*mult_Z + j)/3)*2 < 50)){
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 0]) = tile(((i*mult_Z + j)/3)*2 + 1, 0);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 1]) = tile(((i*mult_Z + j)/3)*2 + 1, 1);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 2]) = tile(((i*mult_Z + j)/3)*2 + 0, 0);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 3]) = tile(((i*mult_Z + j)/3)*2 + 0, 2);
}
else if (((i*mult_Z + j)%3 == 1) && (((i*mult_Z + j)/3)*2 < 50)){
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 0]) = tile(((i*mult_Z + j)/3)*2 + 1, 2 + 0);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 1]) = tile(((i*mult_Z + j)/3)*2 + 1, 3 + 0);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 2]) = tile(((i*mult_Z + j)/3)*2 + 0, 3 + 1);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 3]) = tile(((i*mult_Z + j)/3)*2 + 1, 3 + 1);
}
else {
if (top_x_offset % 5 == 0){
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 0]) = tile(top_x_offset + 0, 5 + 0);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 1]) = tile(top_x_offset + 1, 5 + 0);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 2]) = tile(top_x_offset + 0, 5 + 1);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 3]) = tile(top_x_offset + 0, 5 + 2);
top_x_offset += 2;
}
else if (top_x_offset % 5 == 2){
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 0]) = tile(top_x_offset + 0, 5 + 0);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 1]) = tile(top_x_offset - 1, 5 + 2);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 2]) = tile(top_x_offset + 0, 5 + 2);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 3]) = tile(top_x_offset + 1, 5 + 2);
top_x_offset += 1;
}
else if (top_x_offset % 5 == 3){
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 0]) = tile(top_x_offset + 0, 5 + 0);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 1]) = tile(top_x_offset + 1, 5 + 0);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 2]) = tile(top_x_offset + 0, 5 + 1);
location<kernel>(mat_mul_k[i*mult_Z*mult_Y + j*mult_Y + 3]) = tile(top_x_offset + 1, 5 + 2);
top_x_offset += 2;
}
}
}
}
top_x_offset = 0;
// runtime ratio and place constraints for Add kernels
for (int i = 0; i < mult_X; i++){
for (int j = 0; j < mult_Z; j++){
// for (int k = 0; k < (mult_Y-1); k++){
runtime<ratio>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 0]) = 0.15;
runtime<ratio>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 1]) = 0.15;
runtime<ratio>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 2]) = 0.15;
if (((i*mult_Z + j)%3 == 0) && (((i*mult_Z + j)/3)*2 < 50)){
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 0]) = tile(((i*mult_Z + j)/3)*2 + 0, 1);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 1]) = tile(((i*mult_Z + j)/3)*2 + 0, 1);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 2]) = tile(((i*mult_Z + j)/3)*2 + 0, 1);
}
else if (((i*mult_Z + j)%3 == 1) && (((i*mult_Z + j)/3)*2 < 50)){
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 0]) = tile(((i*mult_Z + j)/3)*2 + 0, 3 + 0);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 1]) = tile(((i*mult_Z + j)/3)*2 + 0, 3 + 0);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 2]) = tile(((i*mult_Z + j)/3)*2 + 0, 3 + 0);
}
else {
if (top_x_offset % 5 == 0){
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 0]) = tile(top_x_offset + 1, 5 + 1);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 1]) = tile(top_x_offset + 1, 5 + 1);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 2]) = tile(top_x_offset + 1, 5 + 1);
top_x_offset += 2;
}
else if (top_x_offset % 5 == 2){
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 0]) = tile(top_x_offset + 0, 5 + 1);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 1]) = tile(top_x_offset + 0, 5 + 1);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 2]) = tile(top_x_offset + 0, 5 + 1);
top_x_offset += 1;
}
else if (top_x_offset % 5 == 3){
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 0]) = tile(top_x_offset + 1, 5 + 1);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 1]) = tile(top_x_offset + 1, 5 + 1);
location<kernel>(add_k[i*mult_Z*(mult_Y-1) + j*(mult_Y-1) + 2]) = tile(top_x_offset + 1, 5 + 1);
top_x_offset += 2;
}
}
// }
}
}
}
};