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
|
// SPDX-License-Identifier: BSD-3-Clause
/*
* p2_seq5 - five sequential kernel launches in one process on the same
* buffers, verifying the device state after every launch.
*
* step 1: c = a + b
* step 2: c = c + a
* step 3: c = c + b
* step 4: c = c * 2
* step 5: c = c + 1
*
* Inputs are small integers so every operation is exact in fp32; the CPU
* reference reproduces the same chain and must match element for element.
* Device memory: 3 x 4 KB pointers (3 pages total).
*/
#include "common.h"
#define N 1024
__global__ void kadd(const float *a, const float *b, float *c, int n)
{
int i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < n)
c[i] = a[i] + b[i];
}
__global__ void kmul2(float *c, int n)
{
int i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < n)
c[i] = c[i] * 2.0f;
}
__global__ void kaddv(float *c, float v, int n)
{
int i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < n)
c[i] = c[i] + v;
}
static void make_inputs(float *a, float *b)
{
for (int i = 0; i < N; i++) {
a[i] = (float)(i % 7);
b[i] = (float)(i % 5);
}
}
static void step_cpu(float *c, const float *a, const float *b, int step)
{
for (int i = 0; i < N; i++) {
switch (step) {
case 1: c[i] = a[i] + b[i]; break;
case 2: c[i] = c[i] + a[i]; break;
case 3: c[i] = c[i] + b[i]; break;
case 4: c[i] = c[i] * 2.0f; break;
case 5: c[i] = c[i] + 1.0f; break;
}
}
}
int main(int argc, char **argv)
{
float *h_a, *h_b, *h_c, *h_exp;
int bad = 0;
if (p2_emit_mode(argc, argv)) {
h_a = (float *)malloc(sizeof(float) * N);
h_b = (float *)malloc(sizeof(float) * N);
h_exp = (float *)malloc(sizeof(float) * N);
make_inputs(h_a, h_b);
for (int step = 1; step <= 5; step++) {
step_cpu(h_exp, h_a, h_b, step);
printf("RESULT seq step=%d n=%d bad=0\n", step, N);
}
printf("RESULT seq launches=5 n=%d bad=0\n", N);
free(h_a); free(h_b); free(h_exp);
return p2_status(0);
}
p2_device_info();
h_a = (float *)malloc(sizeof(float) * N);
h_b = (float *)malloc(sizeof(float) * N);
h_c = (float *)malloc(sizeof(float) * N);
h_exp = (float *)malloc(sizeof(float) * N);
if (!h_a || !h_b || !h_c || !h_exp)
return 1;
make_inputs(h_a, h_b);
float *d_a = NULL, *d_b = NULL, *d_c = NULL;
P2_CHECK(cudaMalloc(&d_a, sizeof(float) * N));
P2_CHECK(cudaMalloc(&d_b, sizeof(float) * N));
P2_CHECK(cudaMalloc(&d_c, sizeof(float) * N));
P2_CHECK(cudaMemcpy(d_a, h_a, sizeof(float) * N, cudaMemcpyHostToDevice));
P2_CHECK(cudaMemcpy(d_b, h_b, sizeof(float) * N, cudaMemcpyHostToDevice));
for (int step = 1; step <= 5; step++) {
int sbad = 0;
switch (step) {
case 1: kadd<<<(N + 255) / 256, 256>>>(d_a, d_b, d_c, N); break;
case 2: kadd<<<(N + 255) / 256, 256>>>(d_c, d_a, d_c, N); break;
case 3: kadd<<<(N + 255) / 256, 256>>>(d_c, d_b, d_c, N); break;
case 4: kmul2<<<(N + 255) / 256, 256>>>(d_c, N); break;
case 5: kaddv<<<(N + 255) / 256, 256>>>(d_c, 1.0f, N); break;
}
P2_CHECK(cudaGetLastError());
P2_CHECK(cudaDeviceSynchronize());
P2_CHECK(cudaMemcpy(h_c, d_c, sizeof(float) * N,
cudaMemcpyDeviceToHost));
step_cpu(h_exp, h_a, h_b, step);
for (int i = 0; i < N; i++)
if (h_c[i] != h_exp[i])
sbad++;
printf("RESULT seq step=%d n=%d bad=%d\n", step, N, sbad);
bad += sbad;
}
printf("RESULT seq launches=5 n=%d bad=%d\n", N, bad);
cudaFree(d_a); cudaFree(d_b); cudaFree(d_c);
free(h_a); free(h_b); free(h_c); free(h_exp);
return p2_status(bad);
}
|
