-
Notifications
You must be signed in to change notification settings - Fork 4
/
stencil-cuda.cu
243 lines (206 loc) · 8.41 KB
/
stencil-cuda.cu
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
#include "prk_util.h"
#include "prk_cuda.h"
__global__ void star2(const int n, const double * in, double * out) {
const int i = blockIdx.y * blockDim.y + threadIdx.y;
const int j = blockIdx.x * blockDim.x + threadIdx.x;
if ( (2 <= i) && (i < n-2) && (2 <= j) && (j < n-2) ) {
out[i*n+j] += +in[(i)*n+(j-2)] * -0.125
+in[(i)*n+(j-1)] * -0.25
+in[(i-2)*n+(j)] * -0.125
+in[(i-1)*n+(j)] * -0.25
+in[(i+1)*n+(j)] * 0.25
+in[(i+2)*n+(j)] * 0.125
+in[(i)*n+(j+1)] * 0.25
+in[(i)*n+(j+2)] * 0.125;
}
}
__global__ void star3(const int n, const double * in, double * out) {
const int i = blockIdx.y * blockDim.y + threadIdx.y;
const int j = blockIdx.x * blockDim.x + threadIdx.x;
if ( (3 <= i) && (i < n-3) && (3 <= j) && (j < n-3) ) {
out[i*n+j] += +in[(i)*n+(j-3)] * -0.05555555555555555
+in[(i)*n+(j-2)] * -0.08333333333333333
+in[(i)*n+(j-1)] * -0.16666666666666666
+in[(i-3)*n+(j)] * -0.05555555555555555
+in[(i-2)*n+(j)] * -0.08333333333333333
+in[(i-1)*n+(j)] * -0.16666666666666666
+in[(i+1)*n+(j)] * 0.16666666666666666
+in[(i+2)*n+(j)] * 0.08333333333333333
+in[(i+3)*n+(j)] * 0.05555555555555555
+in[(i)*n+(j+1)] * 0.16666666666666666
+in[(i)*n+(j+2)] * 0.08333333333333333
+in[(i)*n+(j+3)] * 0.05555555555555555;
}
}
__global__ void star4(const int n, const double * in, double * out) {
const int i = blockIdx.y * blockDim.y + threadIdx.y;
const int j = blockIdx.x * blockDim.x + threadIdx.x;
if ( (4 <= i) && (i < n-4) && (4 <= j) && (j < n-4) ) {
out[i*n+j] += +in[(i)*n+(j-4)] * -0.03125
+in[(i)*n+(j-3)] * -0.041666666666666664
+in[(i)*n+(j-2)] * -0.0625
+in[(i)*n+(j-1)] * -0.125
+in[(i-4)*n+(j)] * -0.03125
+in[(i-3)*n+(j)] * -0.041666666666666664
+in[(i-2)*n+(j)] * -0.0625
+in[(i-1)*n+(j)] * -0.125
+in[(i+1)*n+(j)] * 0.125
+in[(i+2)*n+(j)] * 0.0625
+in[(i+3)*n+(j)] * 0.041666666666666664
+in[(i+4)*n+(j)] * 0.03125
+in[(i)*n+(j+1)] * 0.125
+in[(i)*n+(j+2)] * 0.0625
+in[(i)*n+(j+3)] * 0.041666666666666664
+in[(i)*n+(j+4)] * 0.03125;
}
}
__global__ void nothing(const int n, const double * in, double * out)
{
}
__global__ void add(const int n, double * in)
{
auto i = blockIdx.x * blockDim.x + threadIdx.x;
auto j = blockIdx.y * blockDim.y + threadIdx.y;
if ((i<n) && (j<n)) {
in[i*n+j] += 1.0;
}
}
int main(int argc, char* argv[])
{
std::cout << "Parallel Research Kernels version " << std::endl;
std::cout << "C++11/CUDA Stencil execution on 2D grid" << std::endl;
//////////////////////////////////////////////////////////////////////
// Process and test input parameters
//////////////////////////////////////////////////////////////////////
int iterations;
size_t n, block_size = 16, radius = 2;
try {
if (argc < 3) {
throw "Usage: <# iterations> <array dimension> [<block size> <stencil radius>]";
}
// number of times to run the algorithm
iterations = std::atoi(argv[1]);
if (iterations < 1) {
throw "ERROR: iterations must be >= 1";
}
// linear grid dimension
n = std::atoi(argv[2]);
if (n < 1) {
throw "ERROR: grid dimension must be positive";
} else if (n > prk::get_max_matrix_size()) {
throw "ERROR: grid dimension too large - overflow risk";
}
if (argc > 3) {
block_size = std::atoi(argv[3]);
if (block_size <= 0) block_size = n;
if (block_size > n) block_size = n;
}
if (n % block_size) {
throw "ERROR: block size does not evenly divide grid size";
}
// stencil radius
radius = 2;
if (argc > 4) {
radius = std::atoi(argv[4]);
}
if ( (radius < 1) || (2*radius+1 > n) ) {
throw "ERROR: Stencil radius negative or too large";
}
}
catch (const char * e) {
std::cout << e << std::endl;
return 1;
}
std::cout << "Number of iterations = " << iterations << std::endl;
std::cout << "Grid size = " << n << std::endl;
std::cout << "Block size = " << block_size << std::endl;
std::cout << "Radius of stencil = " << radius << std::endl;
//////////////////////////////////////////////////////////////////////
/// Setup CUDA environment
//////////////////////////////////////////////////////////////////////
prk::CUDA::info info;
info.print(1);
auto stencil = nothing;
switch (radius) {
case 2: stencil = star2; break;
case 3: stencil = star3; break;
case 4: stencil = star4; break;
}
dim3 dimGrid(prk::divceil(n,block_size),prk::divceil(n,block_size),1);
dim3 dimBlock(block_size, block_size, 1);
info.checkDims(dimBlock, dimGrid);
//////////////////////////////////////////////////////////////////////
// Allocate space and perform the computation
//////////////////////////////////////////////////////////////////////
double stencil_time{0};
const size_t nelems = n*n;
const size_t bytes = nelems * sizeof(double);
double * h_in;
double * h_out;
prk::CUDA::check( cudaMallocHost((void**)&h_in, bytes) );
prk::CUDA::check( cudaMallocHost((void**)&h_out, bytes) );
for (int i=0; i<n; i++) {
for (int j=0; j<n; j++) {
h_in[i*n+j] = static_cast<double>(i+j);
h_out[i*n+j] = static_cast<double>(0);
}
}
// copy input from host to device
double * d_in;
double * d_out;
prk::CUDA::check( cudaMalloc((void**)&d_in, bytes) );
prk::CUDA::check( cudaMalloc((void**)&d_out, bytes) );
prk::CUDA::check( cudaMemcpy(d_in, &(h_in[0]), bytes, cudaMemcpyHostToDevice) );
prk::CUDA::check( cudaMemcpy(d_out, &(h_out[0]), bytes, cudaMemcpyHostToDevice) );
prk::CUDA::check( cudaDeviceSynchronize() );
for (int iter = 0; iter<=iterations; iter++) {
if (iter==1) stencil_time = prk::wtime();
// Apply the stencil operator
stencil<<<dimGrid, dimBlock>>>(n, d_in, d_out);
// Add constant to solution to force refresh of neighbor data, if any
add<<<dimGrid, dimBlock>>>(n, d_in);
prk::CUDA::check( cudaDeviceSynchronize() );
}
stencil_time = prk::wtime() - stencil_time;
// copy output back to host
prk::CUDA::check( cudaMemcpy(&(h_out[0]), d_out, bytes, cudaMemcpyDeviceToHost) );
#ifdef VERBOSE
// copy input back to host - debug only
prk::CUDA::check( cudaMemcpy(&(h_in[0]), d_in, bytes, cudaMemcpyDeviceToHost) );
#endif
prk::CUDA::check( cudaFree(d_out) );
prk::CUDA::check( cudaFree(d_in) );
//////////////////////////////////////////////////////////////////////
// Analyze and output results
//////////////////////////////////////////////////////////////////////
// interior of grid with respect to stencil
const size_t active_points = (n-2L*radius)*(n-2L*radius);
double norm{0};
for (size_t i=radius; i<n-radius; i++) {
for (size_t j=radius; j<n-radius; j++) {
norm += prk::abs(h_out[i*n+j]);
}
}
norm /= active_points;
// verify correctness
const double epsilon = 1.0e-8;
const double reference_norm = 2*(iterations+1);
if (prk::abs(norm-reference_norm) > epsilon) {
std::cout << "ERROR: L1 norm = " << norm
<< " Reference L1 norm = " << reference_norm << std::endl;
return 1;
} else {
std::cout << "Solution validates" << std::endl;
#ifdef VERBOSE
std::cout << "L1 norm = " << norm
<< " Reference L1 norm = " << reference_norm << std::endl;
#endif
const size_t stencil_size = 4*radius+1;
size_t flops = (2L*stencil_size+1L) * active_points;
double avgtime = stencil_time/iterations;
std::cout << 8*sizeof(double) << "B "
<< "Rate (MFlops/s): " << 1.0e-6 * static_cast<double>(flops)/avgtime
<< " Avg time (s): " << avgtime << std::endl;
}
return 0;
}