1 Gromacs – OpenCL Porting
10 6. TESTED CONFIGURATIONS
14 - Sharing an OpenCL GPU between two MPI ranks is not supported.
15 See also Issue #91 - https://github.com/StreamComputing/gromacs/issues/91
17 - Using more than one OpenCL GPU on a node is thought to work with thread-MPI,
18 and is known to segfault with AMD OpenCL and OpenMPI on x86 and Linux.
22 - Errors returned by OpenCL functions are handled by using assert calls. This
24 See also Issue #6 - https://github.com/StreamComputing/gromacs/issues/6
26 - clCreateBuffer is always called with CL_MEM_READ_WRITE flag. This needs to be
27 updated so that only the flags that reflect how the buffer is used are provided.
28 For example, if the device is only going to read from a buffer,
29 CL_MEM_READ_ONLY should be used.
30 See also Issue #13 - https://github.com/StreamComputing/gromacs/issues/13
32 - The data structures shared between the OpenCL host and device are defined twice:
33 once in the host code, once in the device code. They must be moved to a single
34 file and shared between the host and the device.
35 See also Issue #16 - https://github.com/StreamComputing/gromacs/issues/16
37 - Quite a few error conditions are unhandled, noted with TODOs in several files
39 - gmx_device_info_t needs struct field documentation
43 - Implement OpenCL kernels for Intel GPUs
45 - Implement OpenCL kernels for Intel CPUs
47 - Improve GPU device sorting in detect_gpus
48 See also Issue #64 - https://github.com/StreamComputing/gromacs/issues/64
50 - Implement warp independent kernels
51 See also Issue #66 - https://github.com/StreamComputing/gromacs/issues/66
53 - Have one OpenCL program object per OpenCL kernel
54 See also Issue #86 - https://github.com/StreamComputing/gromacs/issues/86
56 - Consider parallelising JIT of programs over CPU cores to improve startup
59 - Re-consider caching JIT artefacts to improve startup time
63 - Defining nbparam fields as constants when building the OpenCL kernels
64 See also Issue #87 - https://github.com/StreamComputing/gromacs/issues/87
66 - Fix the tabulated Ewald kernel. This has the potential of being faster than
67 the analytical Ewald kernel
68 See also Issue #65 - https://github.com/StreamComputing/gromacs/issues/65
70 - Evaluate gpu_min_ci_balanced_factor impact on performance for AMD
71 See also Issue #69: https://github.com/StreamComputing/gromacs/issues/69
73 - Update ocl_pmalloc to allocate page locked memory
74 See also Issue #90: https://github.com/StreamComputing/gromacs/issues/90
76 - Update kernel for 128/256threads/block
77 See also Issue #92: https://github.com/StreamComputing/gromacs/issues/92
79 - Update the kernels to use OpenCL 2.0 workgroup level functions if they prove
80 to bring a significant speedup.
81 See also Issue #93: https://github.com/StreamComputing/gromacs/issues/93
83 - Update the kernels to use fixed precision accumulation for force and energy
84 values, if this implementation is faster and does not affect precision.
85 See also Issue #94: https://github.com/StreamComputing/gromacs/issues/94
89 - NVIDIA GPUs are not handled differently depending on compute capability
91 - Because the tabulated kernels have a bug not yet fixed, the current
92 implementation uses only the analytical kernels and never the tabulated ones
93 See also Issue #65 - https://github.com/StreamComputing/gromacs/issues/65
95 - Unlike the CUDA version, the OpenCL implementation uses normal buffers
97 See also Issue #88 - https://github.com/StreamComputing/gromacs/issues/88
99 6. TESTED CONFIGURATIONS
100 =====================
102 NVIDIA GPUs: GeForce GTX 660M, GeForce GTX 750Ti, GeForce GTX 780
103 AMD GPUs: FirePro W5100, HD 7950, FirePro W9100, Radeon R7 M260, R9 290
106 Kernel |Benchmark test |Remarks
107 --------------------------------------------------------------------------------------------------------
108 nbnxn_kernel_ElecCut_VdwLJ_VF_prune_opencl |d.poly-ch2 |
109 nbnxn_kernel_ElecCut_VdwLJ_F_opencl |d.poly-ch2 |
110 nbnxn_kernel_ElecCut_VdwLJ_F_prune_opencl |d.poly-ch2 |
111 nbnxn_kernel_ElecCut_VdwLJ_VF_opencl |d.poly-ch2 |
112 nbnxn_kernel_ElecRF_VdwLJ_VF_prune_opencl |adh_cubic with rf_verlet.mdp |
113 nbnxn_kernel_ElecRF_VdwLJ_F_opencl |adh_cubic with rf_verlet.mdp |
114 nbnxn_kernel_ElecRF_VdwLJ_F_prune_opencl |adh_cubic with rf_verlet.mdp |
115 nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_prune_opencl |adh_cubic_vsites with pme_verlet_vsites.mdp |Failed
116 nbnxn_kernel_ElecEwQSTab_VdwLJ_F_prune_opencl |adh_cubic_vsites with pme_verlet_vsites.mdp |Failed
117 nbnxn_kernel_ElecEw_VdwLJ_VF_prune_opencl |adh_cubic_vsites with pme_verlet_vsites.mdp |
118 nbnxn_kernel_ElecEw_VdwLJ_F_opencl |adh_cubic_vsites with pme_verlet_vsites.mdp |
119 nbnxn_kernel_ElecEw_VdwLJ_F_prune_opencl |adh_cubic_vsites with pme_verlet_vsites.mdp |
120 nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_prune_opencl |adh_cubic_vsites with pme_verlet_vsites.mdp |
121 nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_opencl |adh_cubic_vsites with pme_verlet_vsites.mdp |
123 Input data used for testing - Benchmark data sets available here:
124 ftp://ftp.gromacs.org/pub/benchmarks