1 Gromacs – OpenCL Porting
10 6. TESTED CONFIGURATIONS
14 - Currently there are no known limitations.
18 - Errors returned by OpenCL functions are handled by using assert calls. This
20 See also Issue #6 - https://github.com/StreamComputing/gromacs/issues/6
22 - clCreateBuffer is always called with CL_MEM_READ_WRITE flag. This needs to be
23 updated so that only the flags that reflect how the buffer is used are provided.
24 For example, if the device is only going to read from a buffer,
25 CL_MEM_READ_ONLY should be used.
26 See also Issue #13 - https://github.com/StreamComputing/gromacs/issues/13
28 - The data structures shared between the OpenCL host and device are defined twice:
29 once in the host code, once in the device code. They must be moved to a single
30 file and shared between the host and the device.
31 See also Issue #16 - https://github.com/StreamComputing/gromacs/issues/16
33 - Quite a few error conditions are unhandled, noted with TODOs in several files
35 - gmx_device_info_t needs struct field documentation
39 - Implement OpenCL kernels for Intel GPUs
41 - Implement OpenCL kernels for Intel CPUs
43 - Improve GPU device sorting in detect_gpus
44 See also Issue #64 - https://github.com/StreamComputing/gromacs/issues/64
46 - Implement warp independent kernels
47 See also Issue #66 - https://github.com/StreamComputing/gromacs/issues/66
49 - Have one OpenCL program object per OpenCL kernel
50 See also Issue #86 - https://github.com/StreamComputing/gromacs/issues/86
52 - Consider parallelising JIT of programs over CPU cores to improve startup
55 - Re-consider caching JIT artefacts to improve startup time
59 - Defining nbparam fields as constants when building the OpenCL kernels
60 See also Issue #87 - https://github.com/StreamComputing/gromacs/issues/87
62 - Fix the tabulated Ewald kernel. This has the potential of being faster than
63 the analytical Ewald kernel
64 See also Issue #65 - https://github.com/StreamComputing/gromacs/issues/65
66 - Evaluate gpu_min_ci_balanced_factor impact on performance for AMD
67 See also Issue #69: https://github.com/StreamComputing/gromacs/issues/69
69 - Update ocl_pmalloc to allocate page locked memory
70 See also Issue #90: https://github.com/StreamComputing/gromacs/issues/90
72 - Update kernel for 128/256threads/block
73 See also Issue #92: https://github.com/StreamComputing/gromacs/issues/92
75 - Update the kernels to use OpenCL 2.0 workgroup level functions if they prove
76 to bring a significant speedup.
77 See also Issue #93: https://github.com/StreamComputing/gromacs/issues/93
79 - Update the kernels to use fixed precision accumulation for force and energy
80 values, if this implementation is faster and does not affect precision.
81 See also Issue #94: https://github.com/StreamComputing/gromacs/issues/94
85 - NVIDIA GPUs are not handled differently depending on compute capability
87 - Because the tabulated kernels have a bug not yet fixed, the current
88 implementation uses only the analytical kernels and never the tabulated ones
89 See also Issue #65 - https://github.com/StreamComputing/gromacs/issues/65
91 - Unlike the CUDA version, the OpenCL implementation uses normal buffers
93 See also Issue #88 - https://github.com/StreamComputing/gromacs/issues/88