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
|
#define DOCTEST_CONFIG_IMPLEMENT_WITH_MAIN
#include <doctest.h>
#include <taskflow/taskflow.hpp>
#include <taskflow/cuda/cudaflow.hpp>
template <typename T>
void run_and_wait(T& cf) {
tf::cudaStream stream;
cf.run(stream);
stream.synchronize();
}
// ----------------------------------------------------------------------------
// Matrix Multiplication Kernel
// ----------------------------------------------------------------------------
__global__ void k_multiplication(
int *a, int *b, int *c, int m, int n, int k
) {
int row = blockIdx.y * blockDim.y + threadIdx.y;
int col = blockIdx.x * blockDim.x + threadIdx.x;
int sum = 0;
if(col < k && row < m) {
for(int i = 0; i < n; i++) {
sum += a[row * n + i] * b[i * k + col];
}
c[row * k + col] = sum;
}
}
TEST_CASE("multiply" * doctest::timeout(300)) {
tf::Taskflow taskflow;
tf::Executor executor;
std::vector<int> a, b, c;
const unsigned B = 16;
for(int m=1; m<=256; m<<=1) {
for(int n=1; n<=256; n<<=1) {
for(int k=1; k<=256; k<<=1) {
taskflow.clear();
int* ha {nullptr};
int* hb {nullptr};
int* hc {nullptr};
int* da {nullptr};
int* db {nullptr};
int* dc {nullptr};
dim3 grid ((k+B-1)/B, (m+B-1)/B);
dim3 block (B, B);
auto hosta = taskflow.emplace([&](){
a.resize(m*n);
std::fill_n(a.begin(), m*n, m+n);
ha = a.data();
REQUIRE(cudaMalloc(&da, m*n*sizeof(int)) == cudaSuccess);
}).name("ha");
auto hostb = taskflow.emplace([&](){
b.resize(n*k);
std::fill_n(b.begin(), n*k, n+k);
hb = b.data();
REQUIRE(cudaMalloc(&db, n*k*sizeof(int)) == cudaSuccess);
}).name("hb");
auto hostc = taskflow.emplace([&](){
c.resize(m*k);
hc = c.data();
REQUIRE(cudaMalloc(&dc, m*k*sizeof(int)) == cudaSuccess);
}).name("hc");
auto cuda = taskflow.emplace([&](){
tf::cudaFlow cf;
auto pa = cf.copy(da, ha, m*n);
auto pb = cf.copy(db, hb, n*k);
auto op = cf.kernel(
grid, block, 0, k_multiplication, da, db, dc, m, n, k
).name("op");
auto cc = cf.copy(hc, dc, m*k)
.name("cc");
op.precede(cc).succeed(pa, pb);
run_and_wait(cf);
});
cuda.succeed(hosta, hostb, hostc);
executor.run(taskflow).wait();
for(const auto& x : c) {
REQUIRE(x == (int)(m+n)*(n+k)*n);
}
REQUIRE(cudaFree(da) == cudaSuccess);
REQUIRE(cudaFree(db) == cudaSuccess);
REQUIRE(cudaFree(dc) == cudaSuccess);
}
}
}
}
// ----------------------------------------------------------------------------
// Matrix Transpose
// ----------------------------------------------------------------------------
__global__ void k_transpose(int *mat_in, int *mat_out, int rows, int cols) {
unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
unsigned int idy = blockIdx.y * blockDim.y + threadIdx.y;
if (idx < cols && idy < rows) {
unsigned int pos = idy * cols + idx;
unsigned int trans_pos = idx * rows + idy;
mat_out[trans_pos] = mat_in[pos];
}
}
TEST_CASE("transpose" * doctest::timeout(300)) {
std::vector<int> in, out;
tf::Taskflow taskflow;
tf::Executor executor;
const unsigned B = 16;
for(int m=1; m<=256; m<<=1) {
for(int n=1; n<=256; n<<=1) {
taskflow.clear();
int* ptr_in {nullptr};
int* ptr_out {nullptr};
int* sin {nullptr};
int* sout {nullptr};
dim3 grid ((n+B-1)/B, (m+B-1)/B);
dim3 block (B, B);
auto hin = taskflow.emplace([&](){
in.resize(m*n);
out.resize(m*n);
for(auto& item : in) {
item = ::rand()%100;
}
ptr_in = in.data();
ptr_out = out.data();
REQUIRE(cudaMalloc(&sin, m*n*sizeof(int)) == cudaSuccess);
REQUIRE(cudaMalloc(&sout, m*n*sizeof(int)) == cudaSuccess);
}).name("ha");
auto op = taskflow.emplace([&](){
tf::cudaFlow cf;
auto copyin = cf.copy(sin, ptr_in, m*n);
auto copyout = cf.copy(ptr_out, sout, m*n);
auto trans = cf.kernel(grid, block, 0, k_transpose, sin, sout, m, n);
trans.succeed(copyin).precede(copyout);
run_and_wait(cf);
});
hin.precede(op);
executor.run(taskflow).wait();
for(int x=0; x<m; x++) {
for(int y=0; y<n; ++y) {
REQUIRE(in[x*n+y] == out[y*m+x]);
}
}
REQUIRE(cudaFree(sin) == cudaSuccess);
REQUIRE(cudaFree(sout) == cudaSuccess);
}
}
}
// ----------------------------------------------------------------------------
// vector product
// ----------------------------------------------------------------------------
__global__ void k_product(int *a, int *b, int *c, int N) {
unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < N) {
c[idx] = a[idx] * b[idx];
}
}
TEST_CASE("product" * doctest::timeout(300)) {
tf::Taskflow taskflow;
tf::Executor executor;
const unsigned num_batches = 1024;
const unsigned N = 1024;
const unsigned B = 128;
dim3 grid ((N+B-1)/B);
dim3 block (B);
std::vector<int*> hA(num_batches);
std::vector<int*> hB(num_batches);
std::vector<int*> hC(num_batches);
std::vector<int*> dA(num_batches);
std::vector<int*> dB(num_batches);
std::vector<int*> dC(num_batches);
for(unsigned i=0; i<num_batches; ++i) {
int v1 = ::rand()%10;
int v2 = ::rand()%10;
auto allocate = taskflow.emplace([&, i, v1, v2](){
hA[i] = new int [N];
hB[i] = new int [N];
hC[i] = new int [N];
REQUIRE(cudaMalloc(&dA[i], N*sizeof(int)) == cudaSuccess);
REQUIRE(cudaMalloc(&dB[i], N*sizeof(int)) == cudaSuccess);
REQUIRE(cudaMalloc(&dC[i], N*sizeof(int)) == cudaSuccess);
for(unsigned j=0; j<N; ++j) {
hA[i][j] = v1;
hB[i][j] = v2;
}
});
auto kernel = taskflow.emplace([&, i](){
tf::cudaFlow cf;
auto copyA = cf.copy(dA[i], hA[i], N);
auto copyB = cf.copy(dB[i], hB[i], N);
auto op = cf.kernel(grid, block, 0, k_product, dA[i], dB[i], dC[i], N);
auto copyC = cf.copy(hC[i], dC[i], N);
op.succeed(copyA, copyB).precede(copyC);
run_and_wait(cf);
});
auto deallocate = taskflow.emplace([&, i, v1, v2](){
for(unsigned j=0; j<N; ++j) {
REQUIRE(hC[i][j] == v1*v2);
}
delete hA[i];
delete hB[i];
delete hC[i];
REQUIRE(cudaFree(dA[i]) == cudaSuccess);
REQUIRE(cudaFree(dB[i]) == cudaSuccess);
REQUIRE(cudaFree(dC[i]) == cudaSuccess);
});
kernel.precede(deallocate).succeed(allocate);
}
executor.run(taskflow).wait();
}
|
