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
|
== Sample CUDA application for shared memory bank conflicts ==
Transposes a N x N square matrix of float elements in
global memory and generates an output matrix in global memory.
Defines two versions of CUDA kernel:
transposeCoalesced : Coalesced global memory transpose with shared memory bank conflicts
transposeNoBankConflicts : Coalesced global memory transpose with reduced shared memory bank conflicts
Compiling the code:
==================
> nvcc -lineinfo -gencode arch=compute_70,code=sm_70 -gencode arch=compute_72,code=sm_72 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_87,code=sm_87 -gencode arch=compute_89,code=sm_89 sharedBankConflicts.cu -o sharedBankConflicts
Command line arguments (all are optional):
==========================================
1) <version of kernel to use> Integer value, If not specified uses 1.
1: Use transposeCoalesced() kernel
2: Use transposeNoBankConflicts() kernel
2) <N - Matrix size> Matrix size should be greater than or equal to tile size (TILE_DIM - defined in source file "sharedBankConflicts.cu")
and must be an integral multiple of tile size.
Default value: DEFAULT_MATRIX_SIZE (defined in source file "sharedBankConflicts.cu")
3) <cache config option> String value can be one of "none", "shared", "l1", "equal"
Default value: "none"
Refer the CUDA Runtime API cudaFuncSetCacheConfig() documentation for details of cache configuration.
Sample usage:
============
- Run with default arguments - transposeCoalesced() kernel and default value of N
> ./sharedBankConflicts
- Run with the transposeNoBankConflicts() kernel and default value of N
> ./sharedBankConflicts 2
- Run with the transposeCoalesced() kernel and N=1024
> ./sharedBankConflicts 1 1024
- Run with the transposeNoBankConflicts() kernel with N=1024 and cache config option "l1" (to prefer larger L1 cache and smaller shared memory)
> ./sharedBankConflicts 2 1024 l1
Profiling the sample using Nsight Compute command line
======================================================
- Profile transposeCoalesced() - the initial version of kernel
> ncu --set full --import-source on -o transposeCoalesced.ncu-rep ./sharedBankConflicts 1
- Profile transposeNoBankConflicts() - the updated version of the kernel
> ncu --set full --import-source on -o transposeNoBankConflicts.ncu-rep ./sharedBankConflicts 2
The profiler report files for the sample are also provided and they can be opened in the
Nsight Compute UI using the "File->Open" menu option.
|
