jpekkila
|
b2632c87b4
|
Merge branch 'cmakelist_rewrite_and_C_API_conformity_07-26' into node_device_interface_revision_07-23
|
2019-08-06 15:18:33 +03:00 |
|
jpekkila
|
280804a438
|
Merge branch 'master' into cmakelist_rewrite_and_C_API_conformity_07-26
|
2019-08-06 15:14:33 +03:00 |
|
jpekkila
|
5f4246fb42
|
Standalone now uses O2 optimization level instead of O3. Also removed -march=native since this causes issues if the program is compiled on a different architecture than it is run on. Since we do not do heavy arithmetic on the host side and the host code is not performance-critical part of the code, -march-native is not very useful anyways
|
2019-08-06 14:46:13 +03:00 |
|
jpekkila
|
b73c2675e8
|
Added the optimized implementation of acNodeIntegrate where boundconds are done before integration instead of after
|
2019-08-05 20:10:13 +03:00 |
|
jpekkila
|
8df49370c8
|
Cleanup
|
2019-08-05 19:08:05 +03:00 |
|
jpekkila
|
fa6e1116cb
|
The interface revision now actually works. The issue was incorrect order of src and dst indices when storing the mesh.
|
2019-08-05 17:26:05 +03:00 |
|
jpekkila
|
5232d987c1
|
Added acStoreWithOffset to the revised interface
|
2019-08-05 16:18:22 +03:00 |
|
jpekkila
|
f3de2fa03c
|
Made globalVertexIdx available during preprocessing. NOTE: potentially dangerous. globalVertexIdx should never be used for reading data from the vertex buffers.
|
2019-08-05 15:03:02 +03:00 |
|
jpekkila
|
6dfd03664d
|
Still does not work. I'm starting to think that instead of this one huge revision, we should modify the existing interface step-by-step.
|
2019-08-02 15:31:24 +03:00 |
|
jpekkila
|
5f2378e91b
|
Now compiles (does not work though)
|
2019-08-02 15:15:18 +03:00 |
|
jpekkila
|
567ad61465
|
Multinode MPI implementation should be done later in its own branch. The focus of this branch is to revise the node and device layers. Commented out references to the Grid layer.
|
2019-08-02 13:54:54 +03:00 |
|
jpekkila
|
2b6bf10ae6
|
Dummy implementation of the Grid interface
|
2019-08-01 18:37:36 +03:00 |
|
jpekkila
|
328b809efe
|
Added the revised node interface
|
2019-08-01 14:04:11 +03:00 |
|
jpekkila
|
92376588ba
|
Merge branch 'master' into cmakelist_rewrite_and_C_API_conformity_07-26
|
2019-07-31 20:12:22 +03:00 |
|
jpekkila
|
fb0610c1ba
|
Intermediate changes to the revised node interface
|
2019-07-31 20:04:39 +03:00 |
|
jpekkila
|
0a5d025172
|
Formatting
|
2019-07-31 19:08:16 +03:00 |
|
jpekkila
|
9b7f4277fc
|
Fixed errors in device.cu
|
2019-07-31 19:07:26 +03:00 |
|
jpekkila
|
49026bd26b
|
Revised device interface done
|
2019-07-31 18:46:41 +03:00 |
|
jpekkila
|
5be775dbff
|
Various intermediate changes
|
2019-07-31 17:48:48 +03:00 |
|
jpekkila
|
efd9d54fef
|
Stashing WIP changes (interface revision) s.t. I can continue work on a different machine
|
2019-07-30 14:34:44 +03:00 |
|
jpekkila
|
1ceb6739ae
|
Merge branch 'master' into node_device_interface_revision_07-23
|
2019-07-30 14:31:33 +03:00 |
|
jpekkila
|
62100b1140
|
Merge branch 'master' of https://bitbucket.org/jpekkila/astaroth
|
2019-07-30 14:28:25 +03:00 |
|
jpekkila
|
69deef66fe
|
Added sum reduction. NOTE: Scalar sum does not pass the automated test but vector sum does. I couldn't see anything wrong with the code itself and I strongly suspect that the failures are caused by loss of precision due to summing a huge amount of numbers of different magnitudes. However I'm not yet completely sure. Something like the Kahan summation algorithm might be useful if the errors are really caused by fp arithmetic.
|
2019-07-30 14:28:18 +03:00 |
|
jpekkila
|
fdc1e7333c
|
Added macros for getting int3 and AcReal3 device constants from within kernels (and DSL).
|
2019-07-30 09:10:06 +00:00 |
|
jpekkila
|
c9fafe41e5
|
Tidied the CMakeLists, moved stuff to more logical places and added comments. Also tested that ALTER_CONF=ON still works
|
2019-07-26 15:12:55 +03:00 |
|
jpekkila
|
818893a0ea
|
Fixed stray comma in CUDA_ARCH_FLAGS
|
2019-07-26 14:10:17 +03:00 |
|
jpekkila
|
f322bc8b37
|
Rewrote all CMakeLists. Now much cleaner and there's a clear separation during compilation between the core and standalone modules.
|
2019-07-23 20:50:37 +03:00 |
|
jpekkila
|
b65454d523
|
Stashed some testing files used to make sure that the library can also be used from pure C projects (better compatibility). These changes will never go to master as-is.
|
2019-07-23 18:24:47 +03:00 |
|
jpekkila
|
323d4e3b31
|
Replaced all calls to AC_VTXBUF_IDX to acVertexBufferIdx etc in all files
|
2019-07-23 14:37:28 +03:00 |
|
jpekkila
|
fee03b7149
|
Moved some device limits used only during auto-optimization from astaroth.h to device.cu
|
2019-07-22 19:54:46 +03:00 |
|
jpekkila
|
f74df5339f
|
Cleaned up the include directory: removed all unnecessary stuff and moved common definitions to a separate file
|
2019-07-22 19:46:45 +03:00 |
|
jpekkila
|
01a013f3bc
|
Added WARNCHK_CUDA_ALWAYS to errchk.h
|
2019-07-22 13:05:08 +03:00 |
|
jpekkila
|
a950be99f2
|
Streams now created with priority (all streams have the same priority by default)
|
2019-07-22 13:04:04 +03:00 |
|
jpekkila
|
168b3c4d8b
|
Peer access to neighboring GPUs is now enabled during initialization
|
2019-07-22 13:02:19 +03:00 |
|
jpekkila
|
0db61dd411
|
Disabled the project-wide maxrregcount flag by default since it is only beneficial for resource-heavy kernels. The maximum register count should be defined per kernel instead if needed.
|
2019-07-22 12:58:28 +03:00 |
|
jpekkila
|
78aba6428e
|
Updated the copyright years throughout the project
|
2019-07-16 14:28:32 +03:00 |
|
jpekkila
|
93fc121f5c
|
Introduced versions of the asynchronous functions which take a stream as a parameter
|
2019-07-10 15:49:21 +03:00 |
|
jpekkila
|
bd98eaf9f7
|
Added a stream to loadDeviceConstant call.
|
2019-07-10 15:29:54 +03:00 |
|
jpekkila
|
b08d5b26f5
|
cudaMemcpyToSymbol -> cudaMemcpyToSymbolAsync
|
2019-07-10 15:05:57 +03:00 |
|
jpekkila
|
976bf05c8d
|
Wrong scope for num_iterations in the last commit, fixed
|
2019-07-10 14:37:32 +03:00 |
|
jpekkila
|
866ec8a192
|
Removed some old hack I used for benchmarking a while back
|
2019-07-10 14:34:05 +03:00 |
|
jpekkila
|
d0b95c39b6
|
Disabled writing out unnecessary files when auto-optimizing the code
|
2019-07-09 18:51:04 +03:00 |
|
jpekkila
|
0bda016e17
|
Reviewed the Astaroth interface. Now there's a clear distinction between synchronous and asynchronous functions. For basic usage, we provide a set of functions that are always safe to call (acIntegrate, acLoad, etc), but because of this, must be quite restricted in the sense that f.ex. the whole mesh must be loaded at once and computations cannot be executed concurrently on multiple GPUs. For more advanced users we provide asynchronous functions (such as acLoadWithOffset). Since we cannot know how the asynchronous functions are called (for example, when the integration step has been fully completed and the halos of neighboring subgrids can be safely communicated between GPUs), the responsibility of synchronization must be left to the user. In the existing implementations we currently use only the basic "safe" set of functions (except in renderer.cc), so the existing functionality has not been changed with these latests commits. Autotests also pass.
|
2019-07-09 18:42:00 +03:00 |
|
jpekkila
|
1251f61570
|
Removed a stray acBoundcondStep() in acStore where it definitely shouln't be. Removed code duplication: acBoundcondStep now uses the new acLocalBoundcondStep and acGlobalBoundcondStep functions.
|
2019-07-09 17:08:18 +03:00 |
|
jpekkila
|
10a98b01a9
|
Experimental change: now the integration function is automatically optimized during acInit
|
2019-07-09 14:46:24 +03:00 |
|
jpekkila
|
a086821e7c
|
Added a function acAutoOptimize to the interface and removed rk3_step_async in kernels.cuh (moved into rkStep)
|
2019-07-09 14:21:22 +03:00 |
|
jpekkila
|
84d96de42b
|
Merge branch 'master' into multigpu_optimization_2019-07-05
|
2019-07-09 13:40:33 +03:00 |
|
jpekkila
|
508d15b578
|
Switched from math.h to cmath in math_utils.h. The old-school C math functions are bugged/not overloaded properly in GCC < 6.0 when compiling C++.
|
2019-07-09 13:37:08 +03:00 |
|
jpekkila
|
deebe570da
|
Merge branch 'master' into multigpu_optimization_2019-07-05
|
2019-07-08 16:11:24 +03:00 |
|
Miikka Vaisala
|
6ba15c3a7c
|
props.totalConstMem and props.sharedMemPerBlock cause assembler error
while compiling on TIARA gp cluster. Therefore commeted out.
|
2019-07-08 11:00:12 +08:00 |
|