forked from Yujaeseo/ICDM-2021_MASCOT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsgd_kernel_k64.h
140 lines (114 loc) · 5.44 KB
/
sgd_kernel_k64.h
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
__global__ void single_sgd_k64_hogwild_kernel(
const Node *R,
unsigned int nnz,
float *p,
float *q,
curandState *state,
float lrate,
int k,
int update_count_this_block,
int update_vector_size,
float lambda
)
{
for(int update_ite = 0; update_ite < update_count_this_block; update_ite ++)
{
int lane_id = threadIdx.x%32;
int local_wid = threadIdx.x/32;
int local_w_num = blockDim.x/32;
int wid = local_w_num*blockIdx.x + local_wid;
unsigned int start_id = 0;
if(lane_id == 0)
{
unsigned int origin = (unsigned int)(curand_uniform(&state[wid])*nnz);
start_id = origin%nnz;
}
// All threads read x from laneid 0
start_id = __shfl_sync(0xffffffff,start_id, 0);
for(int i = 0;i < update_vector_size;i++)
{
int offset = (start_id + i)%nnz;
float r = __ldg(&R[offset].r);
int u = __ldg(&R[offset].u);
int v = __ldg(&R[offset].i);
//read the p & q into register file.
int base_p = u*k;
int base_q = v*k;
const float tmp_p1 = p[base_p + lane_id];
const float tmp_q1 = q[base_q + lane_id];
const float tmp_p2 = p[base_p + lane_id + 32];
const float tmp_q2 = q[base_q + lane_id + 32];
float tmp_product = (tmp_p1*tmp_q1) + (tmp_p2*tmp_q2);
tmp_product += __shfl_down_sync(0xffffffff, tmp_product, 16);
tmp_product += __shfl_down_sync(0xffffffff, tmp_product, 8);
tmp_product += __shfl_down_sync(0xffffffff, tmp_product, 4);
tmp_product += __shfl_down_sync(0xffffffff, tmp_product, 2);
tmp_product += __shfl_down_sync(0xffffffff, tmp_product, 1);
tmp_product = __shfl_sync(0xffffffff,tmp_product,0);
float ruv = r - tmp_product;
p[base_p + lane_id + 0] = tmp_p1 + lrate*(ruv*tmp_q1 - lambda*tmp_p1);
q[base_q + lane_id + 0] = tmp_q1 + lrate*(ruv*tmp_p1 - lambda*tmp_q1);
p[base_p + lane_id + 32] = tmp_p2 + lrate*(ruv*tmp_q2 - lambda*tmp_p2);
q[base_q + lane_id + 32] = tmp_q2 + lrate*(ruv*tmp_p2 - lambda*tmp_q2);
}
}
}
__global__ void mem_quant_sgd_k64_hogwild_kernel(
const Node *R,
unsigned int nnz,
half *p,
half *q,
curandState *state,
float lrate,
int k,
int num_iters,
int current_iter,
int update_count_this_block,
int update_vector_size,
float lambda
)
{
for(int update_ite = 0; update_ite < update_count_this_block; update_ite ++)
{
int lane_id = threadIdx.x%32;
int local_wid = threadIdx.x/32;
int local_w_num = blockDim.x/32;
int wid = local_w_num*blockIdx.x + local_wid;
unsigned int start_id = 0;
if(lane_id == 0)
{
unsigned int origin = (unsigned int)(curand_uniform(&state[wid])*nnz);
start_id = origin%nnz;
}
start_id = __shfl_sync(0xffffffff,start_id, 0);
for(int i = 0;i < update_vector_size;i++)
{
int offset = (start_id + i)%nnz;
float r = __ldg(&R[offset].r);
int u = __ldg(&R[offset].u);
int v = __ldg(&R[offset].i);
//read the p & q into register file.
int base_p = u*k;
int base_q = v*k;
float tmp_p1 = __half2float(p[base_p + lane_id]);
float tmp_q1 = __half2float(q[base_q + lane_id]);
float tmp_p2 = __half2float(p[base_p + lane_id + 32]);
float tmp_q2 = __half2float(q[base_q + lane_id + 32]);
float tmp_product = tmp_p1*tmp_q1 + tmp_p2*tmp_q2;
//get dot product.
tmp_product += __shfl_down_sync(0xffffffff,tmp_product, 16);
tmp_product += __shfl_down_sync(0xffffffff,tmp_product, 8);
tmp_product += __shfl_down_sync(0xffffffff,tmp_product, 4);
tmp_product += __shfl_down_sync(0xffffffff,tmp_product, 2);
tmp_product += __shfl_down_sync(0xffffffff,tmp_product, 1);
tmp_product = __shfl_sync(0xffffffff,tmp_product,0);
float ruv = r - tmp_product;
//update
//only works for k=blockDim.x=128
p[base_p + lane_id + 0] = __float2half(tmp_p1 + lrate*(ruv*tmp_q1 - lambda*tmp_p1));
q[base_q + lane_id + 0] = __float2half(tmp_q1 + lrate*(ruv*tmp_p1 - lambda*tmp_q1));
p[base_p + lane_id + 32] = __float2half(tmp_p2 + lrate*(ruv*tmp_q2 - lambda*tmp_p2));
q[base_q + lane_id + 32] = __float2half(tmp_q2 + lrate*(ruv*tmp_p2 - lambda*tmp_q2));
}
}
}