Feature #821
Allow optional 'abort-on-error' instead of exception
| Status: | Resolved | Start date: | 01/23/2012 | |
|---|---|---|---|---|
| Priority: | Normal | Due date: | ||
| Assignee: | M. Rolf | % Done: | 100% | |
| Category: | - | |||
| Target version: | NemoMath 0.4 |
Description
Benchmarked operator[] and *begin() for both read and write operations on revision @191 on malachit with GCC 4.4.3
When throwing exception on 'array out of range':
BenchCase: "VectorCreateOperatorRead"
Estimated cost per operation: 0.00637619 microseconds
BenchCase: "VectorCreateOperatorReadConst"
Estimated cost per operation: 0.000708122 microseconds
BenchCase: "VectorCreateOperatorWrite"
Estimated cost per operation: 0.00803425 microseconds
BenchCase: "VectorCreateIteratorRead"
Estimated cost per operation: 0.00070811 microseconds
BenchCase: "VectorCreateIteratorWrite"
Estimated cost per operation: 0.00283319 microseconds
The same benchmark, when calling 'abort' on 'array out of range':
BenchCase: "VectorCreateOperatorRead"
Estimated cost per operation: 0.000708186 microseconds
BenchCase: "VectorCreateOperatorReadConst"
Estimated cost per operation: 0.000708156 microseconds
BenchCase: "VectorCreateOperatorWrite"
Estimated cost per operation: 0.00283341 microseconds
BenchCase: "VectorCreateIteratorRead"
Estimated cost per operation: 0.00070805 microseconds
BenchCase: "VectorCreateIteratorWrite"
Estimated cost per operation: 0.00283317 microseconds
Non-const operator[] read shows factor 9 speedup, operator[] write factor 3.
In contrast, GCC 4.6.1 shows almost no difference: All costs are on the low level of IteratorRead/Write, except a marginal increase of runtime in "VectorCreateOperatorWrite".
It seems that the older GCC largely benefit from a never-returning error mechanism.
Allow to use abort-on-error using a compiler-flag...
History
#1 Updated by M. Rolf over 11 years ago
- Status changed from New to In Progress
#2 Updated by M. Rolf over 11 years ago
- % Done changed from 0 to 50
#3 Updated by M. Rolf over 11 years ago
- Status changed from In Progress to Resolved
- % Done changed from 50 to 100
Implemented new CMake option ABORT_ON_ERROR (false by default), which activates the compile-time definition -DNEMO_ABORT_ON_ERROR.
When enabled, all exceptions (for example when an index is out of range when accessing a MathVector) are replaced by an error message posted to std::cerr, and an abort()-call which stops the entire program.
Benchmarks with ABORT_ON_ERROR=false (default):
Benchsuite [MathVectorAccess]
Performing each case 1000000000 times with operation-size 10
Benchcase [MathVectorAccess:OperatorRead]
Estimated cost per operation: 0.00699068 us
Benchcase [MathVectorAccess:OperatorReadConst]
Estimated cost per operation: 0.00104675 us
Benchcase [MathVectorAccess:OperatorWrite]
Estimated cost per operation: 0.00959644 us
Benchcase [MathVectorAccess:IteratorRead]
Estimated cost per operation: 0.00070806 us
Benchcase [MathVectorAccess:IteratorWrite]
Estimated cost per operation: 0.00283256 us
Benchsuite [MatrixAccess]
Performing each case 1000000000 times with operation-size 10
Benchcase [MatrixAccess:OperatorRead]
Estimated cost per operation: 0.00596268 us
Benchcase [MatrixAccess:OperatorReadConst]
Estimated cost per operation: 0.00597869 us
Benchcase [MatrixAccess:OperatorWrite]
Estimated cost per operation: 0.00710284 us
Benchcase [MatrixAccess:IteratorRead]
Estimated cost per operation: 0.00105505 us
Benchcase [MatrixAccess:IteratorWrite]
Estimated cost per operation: 0.00292096 us
Benchmarks with ABORT_ON_ERROR=true (some cases x10 faster, due to aggressive optimization of never-returning error-handling):
Benchsuite [MathVectorAccess]
Performing each case 1000000000 times with operation-size 10
Benchcase [MathVectorAccess:OperatorRead]
Estimated cost per operation: 0.000708147 us
Benchcase [MathVectorAccess:OperatorReadConst]
Estimated cost per operation: 0.00106242 us
Benchcase [MathVectorAccess:OperatorWrite]
Estimated cost per operation: 0.00283296 us
Benchcase [MathVectorAccess:IteratorRead]
Estimated cost per operation: 0.000708058 us
Benchcase [MathVectorAccess:IteratorWrite]
Estimated cost per operation: 0.00319325 us
Benchsuite [MatrixAccess]
Performing each case 1000000000 times with operation-size 10
Benchcase [MatrixAccess:OperatorRead]
Estimated cost per operation: 0.000708069 us
Benchcase [MatrixAccess:OperatorReadConst]
Estimated cost per operation: 0.000708512 us
Benchcase [MatrixAccess:OperatorWrite]
Estimated cost per operation: 0.00283275 us
Benchcase [MatrixAccess:IteratorRead]
Estimated cost per operation: 0.00104229 us
Benchcase [MatrixAccess:IteratorWrite]
Estimated cost per operation: 0.00292152 us