Skip to content

Commit

Permalink
updated segscancommon.cu
Browse files Browse the repository at this point in the history
  • Loading branch information
@seanbaxter
seanbaxter committed Sep 29, 2011
1 parent b73bfa1 commit 86f6f2a
Show file tree
Hide file tree
Showing 7 changed files with 1,519 additions and 1,087 deletions.
6 changes: 3 additions & 3 deletions scan/src/mgpuscan/kernelparams.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,9 +14,9 @@
#define PACKED_VALUES_PER_THREAD 16
#define PACKED_BLOCKS_PER_SM 2

#define FLAGS_NUM_THREADS 1024
#define FLAGS_VALUES_PER_THREAD 4
#define FLAGS_BLOCKS_PER_SM 1
#define FLAGS_NUM_THREADS 256
#define FLAGS_VALUES_PER_THREAD 16
#define FLAGS_BLOCKS_PER_SM 2

#define KEYS_NUM_THREADS 256
#define KEYS_VALUES_PER_THREAD 16
Expand Down
Binary file modified scan/src/mgpuscan/mgpuscan.cubin
View file
Binary file not shown.
2,382 changes: 1,387 additions & 995 deletions scan/src/mgpuscan/mgpuscan.isa
View file

Large diffs are not rendered by default.

6 changes: 3 additions & 3 deletions scan/src/mgpuscan/segscancommon.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -89,7 +89,7 @@ DEVICE2 void SegScanDownsweep(uint tid, uint lane, uint warp,
// final thread offsets after the inter-warp multiscan pattern.
uint hasHeadFlag = 0;

/*

if(inclusive) {
#pragma unroll
for(int i = 0; i < ValuesPerThread; ++i) {
Expand All @@ -99,7 +99,7 @@ DEVICE2 void SegScanDownsweep(uint tid, uint lane, uint warp,
x[i] += last;
last = x[i];
}
} else {*/
} else {
#pragma unroll
for(int i = 0; i < ValuesPerThread; ++i) {
if(flags[i]) last = 0;
Expand All @@ -108,7 +108,7 @@ DEVICE2 void SegScanDownsweep(uint tid, uint lane, uint warp,
last += x[i];
x[i] = incLast;
}
// }
}

////////////////////////////////////////////////////////////////////////////
// INTRA-WARP SEGMENT PASS
Expand Down
32 changes: 25 additions & 7 deletions scan/src/scantest/scantest.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,7 @@ const int NumSizes = 3;
const int NumTests = 5;
const int Counts[3] = {
2<< 20,
20<< 20,
10<< 20,
40<< 20
};
const int NumIterations[3] = {
Expand Down Expand Up @@ -147,7 +147,7 @@ bool TestScan(int kind, int count, int numTests, int numIterations,

maxThroughputs[0] = std::max(maxThroughputs[0], mgpuThroughput);

printf("MGPU: %2.6lf B/s\t\t", mgpuThroughput / 1.0e9);
printf("MGPU: %5.2lf B/s\t\t", mgpuThroughput / 1.0e9);
}


Expand Down Expand Up @@ -194,9 +194,9 @@ bool TestScan(int kind, int count, int numTests, int numIterations,
}
#endif // USE_CUDPP

maxThroughputs[1] = std::max(maxThroughputs[0], cudppThroughput);
maxThroughputs[1] = std::max(maxThroughputs[1], cudppThroughput);

printf("CUDPP: %2.6lf B/s\t\t", cudppThroughput / 1.0e9);
printf("CUDPP: %5.2lf B/s\t\t", cudppThroughput / 1.0e9);


////////////////////////////////////////////////////////////////////////
Expand Down Expand Up @@ -230,9 +230,8 @@ bool TestScan(int kind, int count, int numTests, int numIterations,

maxThroughputs[2] = std::max(maxThroughputs[2], thrustThroughput);

printf("thrust: %2.6lf B/s\t\t", thrustThroughput / 1.0e9);
printf("thrust: %5.2lf B/s\t\t", thrustThroughput / 1.0e9);


printf("\n");
}

Expand All @@ -243,6 +242,14 @@ bool TestScan(int kind, int count, int numTests, int numIterations,
return true;
}

void PrintBestTime(const char* label, int test, int kind,
const double throughputs[4][NumSizes][3]) {

printf("%s:\n", label);
for(int i(0); i < NumSizes; ++i)
printf("%5.2lf bn/s\n", throughputs[test][i][kind] / 1.0e9);
}

int main(int argc, char** argv) {


Expand All @@ -268,7 +275,7 @@ int main(int argc, char** argv) {
double throughputs[4][NumSizes][3];
for(int size = 0; size < NumSizes; ++size) {

printf("-------------- %d elements\n", Counts[size]);
printf("\n-------------- %d elements\n", Counts[size]);

printf("Global scan:\n");
TestScan(0, Counts[size], NumTests, NumIterations[size], engine, cudpp,
Expand All @@ -286,5 +293,16 @@ int main(int argc, char** argv) {
TestScan(3, Counts[size], NumTests, NumIterations[size], engine, cudpp,
context, throughputs[3][size]);
}

printf("\nBest times:\n");
PrintBestTime("MGPU scan", 0, 0, throughputs);
PrintBestTime("CUDPP scan", 0, 1, throughputs);
PrintBestTime("thrust scan", 0, 2, throughputs);
PrintBestTime("MGPU seg scan (packed)", 1, 0, throughputs);
PrintBestTime("MGPU seg scan (flags)", 2, 0, throughputs);
PrintBestTime("CUDPP seg scan (flags)", 2, 1, throughputs);
PrintBestTime("MGPU seg scan (keys)", 3, 0, throughputs);
PrintBestTime("thrust seg scan (keys)", 3, 2, throughputs);

}

176 changes: 97 additions & 79 deletions scan/src/timings.txt
View file
Original file line number Diff line number Diff line change
@@ -1,93 +1,111 @@
#define REDUCTION_NUM_THREADS 256

#define SCAN_NUM_THREADS 1024
#define SCAN_VALUES_PER_THREAD 4
#define SCAN_BLOCKS_PER_SM 1

#define PACKED_NUM_THREADS 256
#define PACKED_VALUES_PER_THREAD 16
#define PACKED_BLOCKS_PER_SM 2

#define FLAGS_NUM_THREADS 1024
#define FLAGS_VALUES_PER_THREAD 4
#define FLAGS_BLOCKS_PER_SM 1

#define KEYS_NUM_THREADS 256
#define KEYS_VALUES_PER_THREAD 16
#define KEYS_BLOCKS_PER_SM 2

-------------- 2097152 elements
Global scan:
MGPU: 7.686024 B/s CUDPP: 7.487763 B/s thrust: 5.288093 B/s
MGPU: 7.890411 B/s CUDPP: 7.779953 B/s thrust: 5.246017 B/s
MGPU: 7.747756 B/s CUDPP: 7.795049 B/s thrust: 5.241657 B/s
MGPU: 7.940099 B/s CUDPP: 7.786897 B/s thrust: 5.231989 B/s
MGPU: 7.832367 B/s CUDPP: 7.793913 B/s thrust: 5.246270 B/s
MGPU: 7.59 B/s CUDPP: 7.44 B/s thrust: 5.12 B/s
MGPU: 7.47 B/s CUDPP: 7.73 B/s thrust: 5.29 B/s
MGPU: 7.51 B/s CUDPP: 7.73 B/s thrust: 5.18 B/s
MGPU: 7.79 B/s CUDPP: 7.75 B/s thrust: 5.08 B/s
MGPU: 7.56 B/s CUDPP: 7.80 B/s thrust: 5.20 B/s
Segmented scan (packed):
MGPU: 9.835567 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 9.936244 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 9.941814 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 9.867432 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 9.778741 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 9.64 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 9.46 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 9.48 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 9.91 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 9.64 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
Segmented scan (flags):
MGPU: 7.160979 B/s CUDPP: 7.212241 B/s thrust: 0.000000 B/s
MGPU: 7.370623 B/s CUDPP: 7.220920 B/s thrust: 0.000000 B/s
MGPU: 7.153327 B/s CUDPP: 7.220280 B/s thrust: 0.000000 B/s
MGPU: 7.394118 B/s CUDPP: 7.200216 B/s thrust: 0.000000 B/s
MGPU: 7.234016 B/s CUDPP: 7.173307 B/s thrust: 0.000000 B/s
MGPU: 7.55 B/s CUDPP: 7.23 B/s thrust: 0.00 B/s
MGPU: 7.26 B/s CUDPP: 7.21 B/s thrust: 0.00 B/s
MGPU: 7.36 B/s CUDPP: 7.23 B/s thrust: 0.00 B/s
MGPU: 7.26 B/s CUDPP: 7.22 B/s thrust: 0.00 B/s
MGPU: 7.38 B/s CUDPP: 7.21 B/s thrust: 0.00 B/s
Segmented scan (keys):
MGPU: 7.364980 B/s CUDPP: 0.000000 B/s thrust: 0.863681 B/s
MGPU: 7.353080 B/s CUDPP: 0.000000 B/s thrust: 0.864600 B/s
MGPU: 7.295702 B/s CUDPP: 0.000000 B/s thrust: 0.866747 B/s
MGPU: 7.381601 B/s CUDPP: 0.000000 B/s thrust: 0.866138 B/s
MGPU: 7.248747 B/s CUDPP: 0.000000 B/s thrust: 0.866562 B/s
-------------- 20971520 elements
MGPU: 7.14 B/s CUDPP: 0.00 B/s thrust: 0.86 B/s
MGPU: 7.11 B/s CUDPP: 0.00 B/s thrust: 0.86 B/s
MGPU: 7.21 B/s CUDPP: 0.00 B/s thrust: 0.86 B/s
MGPU: 7.12 B/s CUDPP: 0.00 B/s thrust: 0.87 B/s
MGPU: 6.96 B/s CUDPP: 0.00 B/s thrust: 0.86 B/s

-------------- 10485760 elements
Global scan:
MGPU: 10.467429 B/s CUDPP: 8.344616 B/s thrust: 6.323310 B/s
MGPU: 10.477128 B/s CUDPP: 8.350267 B/s thrust: 6.339872 B/s
MGPU: 10.454303 B/s CUDPP: 8.353797 B/s thrust: 6.221131 B/s
MGPU: 10.443307 B/s CUDPP: 8.347255 B/s thrust: 6.340922 B/s
MGPU: 10.441419 B/s CUDPP: 8.348676 B/s thrust: 6.329775 B/s
MGPU: 9.93 B/s CUDPP: 8.30 B/s thrust: 6.21 B/s
MGPU: 9.97 B/s CUDPP: 8.29 B/s thrust: 6.24 B/s
MGPU: 9.93 B/s CUDPP: 8.30 B/s thrust: 6.26 B/s
MGPU: 10.00 B/s CUDPP: 8.29 B/s thrust: 6.24 B/s
MGPU: 10.00 B/s CUDPP: 8.30 B/s thrust: 6.26 B/s
Segmented scan (packed):
MGPU: 14.554116 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 14.524241 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 14.538165 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 14.490260 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 14.571313 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 13.65 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 13.62 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 13.79 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 13.65 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 13.82 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
Segmented scan (flags):
MGPU: 10.100011 B/s CUDPP: 8.347923 B/s thrust: 0.000000 B/s
MGPU: 10.055328 B/s CUDPP: 8.345563 B/s thrust: 0.000000 B/s
MGPU: 10.076297 B/s CUDPP: 8.347978 B/s thrust: 0.000000 B/s
MGPU: 10.089906 B/s CUDPP: 8.341613 B/s thrust: 0.000000 B/s
MGPU: 10.097860 B/s CUDPP: 8.342602 B/s thrust: 0.000000 B/s
MGPU: 9.65 B/s CUDPP: 8.21 B/s thrust: 0.00 B/s
MGPU: 9.65 B/s CUDPP: 8.21 B/s thrust: 0.00 B/s
MGPU: 9.70 B/s CUDPP: 8.21 B/s thrust: 0.00 B/s
MGPU: 9.56 B/s CUDPP: 8.21 B/s thrust: 0.00 B/s
MGPU: 9.65 B/s CUDPP: 8.22 B/s thrust: 0.00 B/s
Segmented scan (keys):
MGPU: 9.968416 B/s CUDPP: 0.000000 B/s thrust: 1.223831 B/s
MGPU: 9.973149 B/s CUDPP: 0.000000 B/s thrust: 1.224122 B/s
MGPU: 9.969608 B/s CUDPP: 0.000000 B/s thrust: 1.224380 B/s
MGPU: 9.974236 B/s CUDPP: 0.000000 B/s thrust: 1.224434 B/s
MGPU: 10.002642 B/s CUDPP: 0.000000 B/s thrust: 1.224900 B/s
MGPU: 9.54 B/s CUDPP: 0.00 B/s thrust: 1.17 B/s
MGPU: 9.55 B/s CUDPP: 0.00 B/s thrust: 1.17 B/s
MGPU: 9.57 B/s CUDPP: 0.00 B/s thrust: 1.18 B/s
MGPU: 9.59 B/s CUDPP: 0.00 B/s thrust: 1.17 B/s
MGPU: 9.58 B/s CUDPP: 0.00 B/s thrust: 1.17 B/s

-------------- 41943040 elements
Global scan:
MGPU: 10.718733 B/s CUDPP: 8.386371 B/s thrust: 6.360588 B/s
MGPU: 10.708167 B/s CUDPP: 8.385212 B/s thrust: 6.358271 B/s
MGPU: 10.720121 B/s CUDPP: 8.387789 B/s thrust: 6.356108 B/s
MGPU: 10.712956 B/s CUDPP: 8.386780 B/s thrust: 6.362121 B/s
MGPU: 10.719410 B/s CUDPP: 8.385492 B/s thrust: 6.350843 B/s
MGPU: 10.72 B/s CUDPP: 8.41 B/s thrust: 6.42 B/s
MGPU: 10.72 B/s CUDPP: 8.41 B/s thrust: 6.40 B/s
MGPU: 10.72 B/s CUDPP: 8.41 B/s thrust: 6.40 B/s
MGPU: 10.72 B/s CUDPP: 8.41 B/s thrust: 6.40 B/s
MGPU: 10.72 B/s CUDPP: 8.40 B/s thrust: 6.43 B/s
Segmented scan (packed):
MGPU: 15.060673 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 15.054488 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 15.091296 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 15.108144 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 15.116700 B/s CUDPP: 0.000000 B/s thrust: 0.000000 B/s
MGPU: 15.12 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 15.10 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 15.09 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 15.10 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
MGPU: 15.09 B/s CUDPP: 0.00 B/s thrust: 0.00 B/s
Segmented scan (flags):
MGPU: 10.306182 B/s CUDPP: 8.403143 B/s thrust: 0.000000 B/s
MGPU: 10.303040 B/s CUDPP: 8.401961 B/s thrust: 0.000000 B/s
MGPU: 10.303668 B/s CUDPP: 8.402660 B/s thrust: 0.000000 B/s
MGPU: 10.300815 B/s CUDPP: 8.402432 B/s thrust: 0.000000 B/s
MGPU: 10.307584 B/s CUDPP: 8.402082 B/s thrust: 0.000000 B/s
MGPU: 10.31 B/s CUDPP: 8.42 B/s thrust: 0.00 B/s
MGPU: 10.31 B/s CUDPP: 8.42 B/s thrust: 0.00 B/s
MGPU: 10.30 B/s CUDPP: 8.42 B/s thrust: 0.00 B/s
MGPU: 10.33 B/s CUDPP: 8.42 B/s thrust: 0.00 B/s
MGPU: 10.31 B/s CUDPP: 8.42 B/s thrust: 0.00 B/s
Segmented scan (keys):
MGPU: 10.270083 B/s CUDPP: 0.000000 B/s thrust: 1.245405 B/s
MGPU: 10.269416 B/s CUDPP: 0.000000 B/s thrust: 1.246908 B/s
MGPU: 10.269699 B/s CUDPP: 0.000000 B/s thrust: 1.245724 B/s
MGPU: 10.272972 B/s CUDPP: 0.000000 B/s thrust: 1.245341 B/s
MGPU: 10.268508 B/s CUDPP: 0.000000 B/s thrust: 1.245432 B/s
MGPU: 10.28 B/s CUDPP: 0.00 B/s thrust: 1.24 B/s
MGPU: 10.28 B/s CUDPP: 0.00 B/s thrust: 1.24 B/s
MGPU: 10.27 B/s CUDPP: 0.00 B/s thrust: 1.24 B/s
MGPU: 10.28 B/s CUDPP: 0.00 B/s thrust: 1.24 B/s
MGPU: 10.27 B/s CUDPP: 0.00 B/s thrust: 1.24 B/s

Best times:
MGPU scan:
7.79 bn/s
10.00 bn/s
10.72 bn/s
CUDPP scan:
7.80 bn/s
8.30 bn/s
8.41 bn/s
thrust scan:
5.29 bn/s
6.26 bn/s
6.43 bn/s
MGPU seg scan (packed):
9.91 bn/s
13.82 bn/s
15.12 bn/s
MGPU seg scan (flags):
7.55 bn/s
9.70 bn/s
10.33 bn/s
CUDPP seg scan (flags):
7.23 bn/s
8.22 bn/s
8.42 bn/s
MGPU seg scan (keys):
7.21 bn/s
9.59 bn/s
10.28 bn/s
thrust seg scan (keys):
0.87 bn/s
1.18 bn/s
1.24 bn/s
4 changes: 4 additions & 0 deletions scan/vs9/mgpuscan/mgpuscan.vcproj
View file
Original file line number Diff line number Diff line change
Expand Up @@ -219,6 +219,10 @@
RelativePath="..\..\src\mgpuscan\globalscan.cu"
>
</File>
<File
RelativePath="..\..\src\mgpuscan\loadstore.cu"
>
</File>
<File
RelativePath="..\..\src\mgpuscan\scancommon.cu"
>
Expand Down

0 comments on commit 86f6f2a

Please sign in to comment.