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
|
#include <cuda.h>
#include <stdint.h>
#include "config.h"
#include "ipedma.h"
__global__ void null(uint32_t *data) {
}
__device__ void ksleep(uint32_t clocks) {
clock_t start = clock64(), now;
do {
now = clock64();
} while ((start < now)&&((now - start) < clocks));
}
__global__ void ipedma(volatile void *bar, uintptr_t bus_addr, volatile uint32_t *counter, volatile uint64_t *desc, uint32_t *data) {
int i, j;
clock_t sum = 0, sumk = 0, t1, t2, t3;
for (i = 0; i < GPU_ITERS; i++) {
long wait = 0;
// It would not work as we don't know in which order threads/blocks are executed. We also can't push 0 in all threads
// as we unsure if it will overwrite. In non-iterative use case we do not need to push zero and it could work.
// Single thread of block should poll and, then, we synchronize. Limiting amount of blocks will be good...
if ((threadIdx.x == 0)&&(blockIdx.x == 0)) {
// Reset counter
//desc[1] = 0;
*counter = 0;
#ifdef USE_HW_CONTER
WR32 (REG_DMA, 0);
WR32 (REG_PERF_COUNTER, 0);
for (j = 0; j < NUM_PAGES; j++) {
WR64 (REG_DESCRIPTOR_ADDRESS, bus_addr);
// ksleep(10000);
}
t1 = clock64();
WR32 (REG_DMA, 1);
#else
t1 = clock64();
for (j = 0; j < NUM_PAGES; j++) {
WR64 (REG_DESCRIPTOR_ADDRESS, bus_addr);
// ksleep(10000);
}
#endif
}
do {
if (++wait > 0x10000) break;
} while (((*counter) < (NUM_PAGES))/*||(desc[1] == 0)*/);
t2 = clock64();
#ifdef USE_HW_CONTER
WR32 (REG_PERF_COUNTER, 1);
#endif
null<<<1,1>>>(data);
cudaDeviceSynchronize();
t3 = clock64();
sum += t2 - t1;
sumk += t3 - t1;
}
if ((threadIdx.x == 0)&&(blockIdx.x == 0)) {
data[0] = sum / GPU_ITERS;
data[1] = sumk / GPU_ITERS;
data[2] = *counter;
}
}
/*
__global__ void do_leet (int *rin) {
*rin = 0x1337;
}
*/
|
