1 | #include "NBody.hpp" |
---|
2 | |
---|
3 | extern int numBodies; |
---|
4 | |
---|
5 | /////////////////////////////////// NBody::setupCLKernels Func \\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\ |
---|
6 | // |
---|
7 | // This function only sets the Kernel arguments |
---|
8 | // |
---|
9 | /////////////////////////////////// NBody::setupCLKernels Func \\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\ |
---|
10 | |
---|
11 | int NBody::setupCLKernels() |
---|
12 | { |
---|
13 | cl_int status; |
---|
14 | |
---|
15 | /* get a kernel object handle for a kernel with the given name */ |
---|
16 | kernel = clCreateKernel( |
---|
17 | program, |
---|
18 | kernelFunctionName, |
---|
19 | &status); |
---|
20 | |
---|
21 | if(status != CL_SUCCESS) |
---|
22 | { |
---|
23 | std::cout << "clCreateKernel failed." << std::endl; |
---|
24 | return 1; |
---|
25 | } |
---|
26 | |
---|
27 | //================================================ |
---|
28 | // Arguments specific to all kernels: |
---|
29 | |
---|
30 | /*** Set appropriate arguments to the kernel ***/ |
---|
31 | |
---|
32 | /* Particle positions */ |
---|
33 | status = clSetKernelArg( |
---|
34 | kernel, |
---|
35 | 0, // Argument one; |
---|
36 | sizeof(cl_mem), |
---|
37 | (void *)&updatedPos); |
---|
38 | |
---|
39 | if(status != CL_SUCCESS) |
---|
40 | { |
---|
41 | std::cout << "clSetKernelArg failed. (updatedPos)" << std::endl; |
---|
42 | return 1; |
---|
43 | } |
---|
44 | |
---|
45 | /* Particle velocity */ |
---|
46 | status = clSetKernelArg( |
---|
47 | kernel, |
---|
48 | 1, // Argument two; |
---|
49 | sizeof(cl_mem), |
---|
50 | (void *)&updatedVel); |
---|
51 | |
---|
52 | if(status != CL_SUCCESS) |
---|
53 | { |
---|
54 | std::cout << "clSetKernelArg failed. (updatedVel)" << std::endl; |
---|
55 | return 1; |
---|
56 | } |
---|
57 | |
---|
58 | /* numBodies */ |
---|
59 | status = clSetKernelArg( |
---|
60 | kernel, |
---|
61 | 2, // Argument three; |
---|
62 | sizeof(cl_int), |
---|
63 | (void *)&numBodies); |
---|
64 | |
---|
65 | if(status != CL_SUCCESS) |
---|
66 | { |
---|
67 | std::cout << "clSetKernelArg failed. (numBodies)" << std::endl; |
---|
68 | return 1; |
---|
69 | } |
---|
70 | |
---|
71 | /* time step */ |
---|
72 | status = clSetKernelArg( |
---|
73 | kernel, |
---|
74 | 3, // Argument four; |
---|
75 | sizeof(cl_float), |
---|
76 | (void *)&delT); |
---|
77 | if(status != CL_SUCCESS) |
---|
78 | { |
---|
79 | std::cout << "clSetKernelArg failed. (delT)" << std::endl; |
---|
80 | return 1; |
---|
81 | } |
---|
82 | |
---|
83 | /* upward Pseudoprobability */ |
---|
84 | status = clSetKernelArg( |
---|
85 | kernel, |
---|
86 | 4, // Argument five; |
---|
87 | sizeof(cl_float), |
---|
88 | (void *)&espSqr); |
---|
89 | |
---|
90 | if(status != CL_SUCCESS) |
---|
91 | { |
---|
92 | std::cout << "clSetKernelArg failed. (espSqr)" << std::endl; |
---|
93 | return 1; |
---|
94 | } |
---|
95 | |
---|
96 | /* local memory */ |
---|
97 | status = clSetKernelArg( |
---|
98 | kernel, |
---|
99 | 5, // Argument six; |
---|
100 | GROUP_SIZE * 4 * sizeof(float), |
---|
101 | NULL); |
---|
102 | |
---|
103 | if(status != CL_SUCCESS) |
---|
104 | { |
---|
105 | std::cout << "clSetKernelArg failed. (localPos)" << std::endl; |
---|
106 | return 1; |
---|
107 | } |
---|
108 | |
---|
109 | //================================================ |
---|
110 | // Arguments specific to Leapfrog kernel: |
---|
111 | |
---|
112 | if (integrator == LEAPFROG){ |
---|
113 | |
---|
114 | /* local memory */ |
---|
115 | status = clSetKernelArg( |
---|
116 | kernel, |
---|
117 | 6, // Argument seven; |
---|
118 | GROUP_SIZE * 4 * sizeof(float), |
---|
119 | NULL); |
---|
120 | |
---|
121 | if(status != CL_SUCCESS) |
---|
122 | { |
---|
123 | std::cout << "clSetKernelArg failed. (localVel)" << std::endl; |
---|
124 | return 1; |
---|
125 | } |
---|
126 | |
---|
127 | /* Particle acceleration */ |
---|
128 | status = clSetKernelArg( |
---|
129 | kernel, |
---|
130 | 7, // Argument eight; |
---|
131 | sizeof(cl_mem), |
---|
132 | (void *)&updatedAcc); |
---|
133 | |
---|
134 | if(status != CL_SUCCESS) |
---|
135 | { |
---|
136 | std::cout << "clSetKernelArg failed. (updatedAcc)" << std::endl; |
---|
137 | return 1; |
---|
138 | } |
---|
139 | |
---|
140 | /* If it is the first run and we do not have the acc computed yet */ |
---|
141 | status = clSetKernelArg( |
---|
142 | kernel, |
---|
143 | 8, // Argument nine; |
---|
144 | sizeof(cl_int), |
---|
145 | (void *)&initialAccComputation); |
---|
146 | |
---|
147 | if(status != CL_SUCCESS) |
---|
148 | { |
---|
149 | std::cout << "clSetKernelArg failed. (initialAccComputation)" << std::endl; |
---|
150 | return 1; |
---|
151 | } |
---|
152 | |
---|
153 | /* If it is the first run and we do not have the acc computed yet */ |
---|
154 | status = clSetKernelArg( |
---|
155 | kernel, |
---|
156 | 9, // Argument 10; |
---|
157 | sizeof(cl_mem), |
---|
158 | (void *)&updatedCollTime); |
---|
159 | |
---|
160 | if(status != CL_SUCCESS) |
---|
161 | { |
---|
162 | std::cout << "clSetKernelArg failed. (updatedCollTime)" << std::endl; |
---|
163 | return 1; |
---|
164 | } |
---|
165 | } |
---|
166 | |
---|
167 | //================================================ |
---|
168 | // Try to see if there is enough local memory |
---|
169 | // on the device for the above parameter |
---|
170 | // declared to be __local: |
---|
171 | status = clGetKernelWorkGroupInfo(kernel, |
---|
172 | devices[0], |
---|
173 | CL_KERNEL_LOCAL_MEM_SIZE, |
---|
174 | sizeof(cl_ulong), |
---|
175 | &usedLocalMemory, |
---|
176 | NULL); |
---|
177 | if(status != CL_SUCCESS) |
---|
178 | { |
---|
179 | std::cout << "clGetKernelWorkGroupInfo CL_KERNEL_LOCAL_MEM_SIZE failed." << std::endl; |
---|
180 | return 1; |
---|
181 | } |
---|
182 | |
---|
183 | if(usedLocalMemory > totalLocalMemory) |
---|
184 | { |
---|
185 | std::cout << "Unsupported: Insufficient local memory on device." << std::endl; |
---|
186 | return 1; |
---|
187 | } |
---|
188 | |
---|
189 | return 0; |
---|
190 | } |
---|