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
|
/**
* app.c
* SpMV Host Application Source File
*
*/
#include <dpu.h>
#include <dpu_log.h>
#include <assert.h>
#include <getopt.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include "mram-management.h"
#include "../support/common.h"
#include "../support/matrix.h"
#include "../support/params.h"
#include "../support/timer.h"
#include "../support/utils.h"
#define DPU_BINARY "./bin/dpu_code"
#define XSTR(x) STR(x)
#define STR(x) #x
#ifndef ENERGY
#define ENERGY 0
#endif
#if ENERGY
#include <dpu_probe.h>
#endif
// Main of the Host Application
int main(int argc, char** argv) {
// Process parameters
struct Params p = input_params(argc, argv);
// Timing and profiling
Timer timer;
double allocTime = 0.0f, loadTime = 0.0f, writeTime = 0.0f, dpuTime = 0.0f, readTime = 0.0f, freeTime = 0.0f;
#if ENERGY
struct dpu_probe_t probe;
DPU_ASSERT(dpu_probe_init("energy_probe", &probe));
#endif
// Allocate DPUs and load binary
struct dpu_set_t dpu_set, dpu;
uint32_t numDPUs, numRanks;
startTimer(&timer);
DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set));
stopTimer(&timer);
allocTime += getElapsedTime(timer);
startTimer(&timer);
DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL));
stopTimer(&timer);
loadTime += getElapsedTime(timer);
DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &numDPUs));
DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &numRanks));
assert(numDPUs == NR_DPUS);
PRINT_INFO(p.verbosity >= 1, "Allocated %d DPU(s)", numDPUs);
// Initialize SpMV data structures
PRINT_INFO(p.verbosity >= 1, "Reading matrix %s", p.fileName);
struct COOMatrix cooMatrix = readCOOMatrix(p.fileName);
PRINT_INFO(p.verbosity >= 1, " %u rows, %u columns, %u nonzeros", cooMatrix.numRows, cooMatrix.numCols, cooMatrix.numNonzeros);
struct CSRMatrix csrMatrix = coo2csr(cooMatrix);
uint32_t numRows = csrMatrix.numRows;
uint32_t numCols = csrMatrix.numCols;
uint32_t* rowPtrs = csrMatrix.rowPtrs;
struct Nonzero* nonzeros = csrMatrix.nonzeros;
float* inVector = malloc(ROUND_UP_TO_MULTIPLE_OF_8(numCols*sizeof(float)));
initVector(inVector, numCols);
float* outVector = malloc(ROUND_UP_TO_MULTIPLE_OF_8(numRows*sizeof(float)));
// Partition data structure across DPUs
uint32_t numRowsPerDPU = ROUND_UP_TO_MULTIPLE_OF_2((numRows - 1)/numDPUs + 1);
PRINT_INFO(p.verbosity >= 1, "Assigning %u rows per DPU", numRowsPerDPU);
struct DPUParams dpuParams[numDPUs];
unsigned int dpuIdx = 0;
PRINT_INFO(p.verbosity == 1, "Copying data to DPUs");
DPU_FOREACH (dpu_set, dpu) {
// Allocate parameters
struct mram_heap_allocator_t allocator;
init_allocator(&allocator);
uint32_t dpuParams_m = mram_heap_alloc(&allocator, sizeof(struct DPUParams));
// Find DPU's rows
uint32_t dpuStartRowIdx = dpuIdx*numRowsPerDPU;
uint32_t dpuNumRows;
if(dpuStartRowIdx > numRows) {
dpuNumRows = 0;
} else if(dpuStartRowIdx + numRowsPerDPU > numRows) {
dpuNumRows = numRows - dpuStartRowIdx;
} else {
dpuNumRows = numRowsPerDPU;
}
dpuParams[dpuIdx].dpuNumRows = dpuNumRows;
PRINT_INFO(p.verbosity >= 2, " DPU %u:", dpuIdx);
PRINT_INFO(p.verbosity >= 2, " Receives %u rows", dpuNumRows);
// Partition nonzeros and copy data
if(dpuNumRows > 0) {
// Find DPU's CSR matrix partition
uint32_t* dpuRowPtrs_h = &rowPtrs[dpuStartRowIdx];
uint32_t dpuRowPtrsOffset = dpuRowPtrs_h[0];
struct Nonzero* dpuNonzeros_h = &nonzeros[dpuRowPtrsOffset];
uint32_t dpuNumNonzeros = dpuRowPtrs_h[dpuNumRows] - dpuRowPtrsOffset;
// Allocate MRAM
uint32_t dpuRowPtrs_m = mram_heap_alloc(&allocator, (dpuNumRows + 1)*sizeof(uint32_t));
uint32_t dpuNonzeros_m = mram_heap_alloc(&allocator, dpuNumNonzeros*sizeof(struct Nonzero));
uint32_t dpuInVector_m = mram_heap_alloc(&allocator, numCols*sizeof(float));
uint32_t dpuOutVector_m = mram_heap_alloc(&allocator, dpuNumRows*sizeof(float));
assert((dpuNumRows*sizeof(float))%8 == 0 && "Output sub-vector must be a multiple of 8 bytes!");
PRINT_INFO(p.verbosity >= 2, " Total memory allocated is %d bytes", allocator.totalAllocated);
// Set up DPU parameters
dpuParams[dpuIdx].dpuRowPtrsOffset = dpuRowPtrsOffset;
dpuParams[dpuIdx].dpuRowPtrs_m = dpuRowPtrs_m;
dpuParams[dpuIdx].dpuNonzeros_m = dpuNonzeros_m;
dpuParams[dpuIdx].dpuInVector_m = dpuInVector_m;
dpuParams[dpuIdx].dpuOutVector_m = dpuOutVector_m;
// Send data to DPU
PRINT_INFO(p.verbosity >= 2, " Copying data to DPU");
startTimer(&timer);
copyToDPU(dpu, (uint8_t*)dpuRowPtrs_h, dpuRowPtrs_m, (dpuNumRows + 1)*sizeof(uint32_t));
copyToDPU(dpu, (uint8_t*)dpuNonzeros_h, dpuNonzeros_m, dpuNumNonzeros*sizeof(struct Nonzero));
copyToDPU(dpu, (uint8_t*)inVector, dpuInVector_m, numCols*sizeof(float));
stopTimer(&timer);
writeTime += getElapsedTime(timer);
}
// Send parameters to DPU
PRINT_INFO(p.verbosity >= 2, " Copying parameters to DPU");
startTimer(&timer);
copyToDPU(dpu, (uint8_t*)&dpuParams[dpuIdx], dpuParams_m, sizeof(struct DPUParams));
stopTimer(&timer);
writeTime += getElapsedTime(timer);
++dpuIdx;
}
PRINT_INFO(p.verbosity >= 1, " CPU-DPU Time: %f ms", writeTime*1e3);
// Run all DPUs
PRINT_INFO(p.verbosity >= 1, "Booting DPUs");
startTimer(&timer);
#if ENERGY
DPU_ASSERT(dpu_probe_start(&probe));
#endif
DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS));
#if ENERGY
DPU_ASSERT(dpu_probe_stop(&probe));
double energy;
DPU_ASSERT(dpu_probe_get(&probe, DPU_ENERGY, DPU_AVERAGE, &energy));
PRINT_INFO(p.verbosity >= 1, " DPU Energy: %f J", energy);
#endif
stopTimer(&timer);
dpuTime += getElapsedTime(timer);
PRINT_INFO(p.verbosity >= 1, " DPU Time: %f ms", dpuTime*1e3);
// Copy back result
PRINT_INFO(p.verbosity >= 1, "Copying back the result");
startTimer(&timer);
dpuIdx = 0;
DPU_FOREACH (dpu_set, dpu) {
unsigned int dpuNumRows = dpuParams[dpuIdx].dpuNumRows;
if(dpuNumRows > 0) {
uint32_t dpuStartRowIdx = dpuIdx*numRowsPerDPU;
copyFromDPU(dpu, dpuParams[dpuIdx].dpuOutVector_m, (uint8_t*)(outVector + dpuStartRowIdx), dpuNumRows*sizeof(float));
}
++dpuIdx;
}
stopTimer(&timer);
readTime += getElapsedTime(timer);
PRINT_INFO(p.verbosity >= 1, " DPU-CPU Time: %f ms", readTime*1e3);
// Calculating result on CPU
PRINT_INFO(p.verbosity >= 1, "Calculating result on CPU");
float* outVectorReference = malloc(numRows*sizeof(float));
for(uint32_t rowIdx = 0; rowIdx < numRows; ++rowIdx) {
float sum = 0.0f;
for(uint32_t i = rowPtrs[rowIdx]; i < rowPtrs[rowIdx + 1]; ++i) {
uint32_t colIdx = nonzeros[i].col;
float value = nonzeros[i].value;
sum += inVector[colIdx]*value;
}
outVectorReference[rowIdx] = sum;
}
// Verify the result
PRINT_INFO(p.verbosity >= 1, "Verifying the result");
int status = 1;
for(uint32_t rowIdx = 0; rowIdx < numRows; ++rowIdx) {
float diff = (outVectorReference[rowIdx] - outVector[rowIdx])/outVectorReference[rowIdx];
const float tolerance = 0.00001;
if(diff > tolerance || diff < -tolerance) {
status = 0;
PRINT_ERROR("Mismatch at index %u (CPU result = %f, DPU result = %f)", rowIdx, outVectorReference[rowIdx], outVector[rowIdx]);
}
}
startTimer(&timer);
DPU_ASSERT(dpu_free(dpu_set));
stopTimer(&timer);
freeTime += getElapsedTime(timer);
if (status) {
printf("[::] SpMV UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s n_elements=%d ",
numDPUs, numRanks, NR_TASKLETS, "float", csrMatrix.numNonzeros);
printf("| latency_alloc_us=%f latency_load_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f latency_free_us=%f",
allocTime, loadTime, writeTime, dpuTime, readTime, freeTime);
printf(" throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f",
// coomatrix / csrmatrix use uint32_t indexes and float values, so all 32bit
csrMatrix.numNonzeros * sizeof(float) / (dpuTime * 1e6),
csrMatrix.numNonzeros * sizeof(float) / ((allocTime + loadTime + writeTime + dpuTime + readTime + freeTime) * 1e6));
printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f",
csrMatrix.numNonzeros * sizeof(float) / ((writeTime + dpuTime + readTime) * 1e6),
csrMatrix.numNonzeros * sizeof(float) / ((loadTime + writeTime + dpuTime + readTime) * 1e6),
csrMatrix.numNonzeros * sizeof(float) / ((allocTime + loadTime + writeTime + dpuTime + readTime) * 1e6));
printf(" throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f",
// coomatrix / csrmatrix use uint32_t indexes and float values, so all 32bit
csrMatrix.numNonzeros / (dpuTime * 1e6),
csrMatrix.numNonzeros / ((allocTime + loadTime + writeTime + dpuTime + readTime + freeTime) * 1e6));
printf(" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f",
csrMatrix.numNonzeros / ((writeTime + dpuTime + readTime) * 1e6),
csrMatrix.numNonzeros / ((loadTime + writeTime + dpuTime + readTime) * 1e6),
csrMatrix.numNonzeros / ((allocTime + loadTime + writeTime + dpuTime + readTime) * 1e6));
}
// Display DPU Logs
if(p.verbosity >= 2) {
PRINT_INFO(p.verbosity >= 2, "Displaying DPU Logs:");
dpuIdx = 0;
DPU_FOREACH (dpu_set, dpu) {
PRINT("DPU %u:", dpuIdx);
DPU_ASSERT(dpu_log_read(dpu, stdout));
++dpuIdx;
}
}
// Deallocate data structures
freeCOOMatrix(cooMatrix);
freeCSRMatrix(csrMatrix);
free(inVector);
free(outVector);
free(outVectorReference);
return 0;
}
|
