Bytecode for transfer/Evaluation - Revision history

Unhammer: cat

2013-09-24T06:52:03Z

cat

Jacob Nordfalk at 20:13, 13 October 2010

2010-10-13T20:13:53Z

Jacob Nordfalk: moved Bytecode for transfer/Evaluation on all released language pairs to Bytecode for transfer/Evaluation: Shorter URL

2010-10-11T10:24:22Z

moved Bytecode for transfer/Evaluation on all released language pairs to Bytecode for transfer/Evaluation: Shorter URL

85.233.243.36 at 04:35, 11 October 2010

2010-10-11T04:35:52Z

Jacob Nordfalk at 02:41, 28 February 2010

2010-02-28T02:41:30Z

Jacob Nordfalk at 02:40, 28 February 2010

2010-02-28T02:40:12Z

Jacob Nordfalk at 02:12, 28 February 2010

2010-02-28T02:12:22Z

Jacob Nordfalk at 01:34, 28 February 2010

2010-02-28T01:34:00Z

Jacob Nordfalk at 01:11, 28 February 2010

2010-02-28T01:11:10Z

Jacob Nordfalk at 01:09, 28 February 2010

2010-02-28T01:09:57Z

← Older revision		Revision as of 06:52, 24 September 2013
Line 793:		Line 793:

	</pre>		</pre>
			[[Category:Transfer]]

@@ Line 8: / Line 8: @@
 The individual numbers can be seen in the printout below: 4.36, 4.29, 5.04, 4.49, 5.76, 4.29, 5.06, 4.92, 4.05, 5.37, 5.02, 5.63, 6.87, 5.83, 5.78, 4.64, 4.76, 5.45, 4.6, 5.21, 4.06, 5.32, 4.5, 5.76, 4.67, 5.41, 5.68, 6.63, 5.37, 6.13, 4.84, 5.6, 3.82, 4.7, 5.13, 4.92, 6.12, 4.72, 3.95, 5.33, 5.3, 4.62, 4.01, 3.21, 7.8, 3.48, 4.49, 4.97, 4.73, 4.78, 3.81, 4.74, 4.62, 5.9, 4.19, 4.55, 3.2, 4.05, 4.32, 5.64, 5.03, 5.11.
+As the other stages also take some time this usually this gives a speed increase of a factor 3 in the total translation time.
 === Validation checks ===

@@ Line 3: / Line 3: @@
 It is done by running the class [http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/lttoolbox-java/src/org/apertium/transfer/development/FindAndCompareAllReleasedTransferFiles.java?revision=20387&view=markup FindAndCompareAllReleasedTransferFiles]
+=== Average speedup ===
+The average speedup of running bytecode transfer is around 5 (4.946) with very few with a speedup of more than 6 or speedup less than 4.
+The individual numbers can be seen in the printout below: 4.36, 4.29, 5.04, 4.49, 5.76, 4.29, 5.06, 4.92, 4.05, 5.37, 5.02, 5.63, 6.87, 5.83, 5.78, 4.64, 4.76, 5.45, 4.6, 5.21, 4.06, 5.32, 4.5, 5.76, 4.67, 5.41, 5.68, 6.63, 5.37, 6.13, 4.84, 5.6, 3.82, 4.7, 5.13, 4.92, 6.12, 4.72, 3.95, 5.33, 5.3, 4.62, 4.01, 3.21, 7.8, 3.48, 4.49, 4.97, 4.73, 4.78, 3.81, 4.74, 4.62, 5.9, 4.19, 4.55, 3.2, 4.05, 4.32, 5.64, 5.03, 5.11.
+=== Validation checks ===
 First each file is compiled to Java bytecode. A lot of pairs contains errors which cant be found by apertium-validate-transfer, but which is found by the transfer-to-bytecode compiler. This includes:

← Older revision		Revision as of 02:41, 28 February 2010
Line 1:		Line 1:
	This is a test of all .t1x files in all language pairs in http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/. There is 63 of them, and '''in all cases the bytecode transfer gives exactly the same result''' as the old interpreted transfer, down to the handling of wrong clip statements (but excluding invoking macros with wrong ~~numer of parameters or~~ parameters, as the results of that ~~are~~ unpredictable).		This is a test of all .t1x files in all language pairs in http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/. There is 63 of them, and '''in all cases the bytecode transfer gives exactly the same result''' as the old interpreted transfer, down to the handling of wrong clip statements (but excluding invoking macros with wrong parameters, as the results of that is unpredictable memory reads).

	It is done by running the class [http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/lttoolbox-java/src/org/apertium/transfer/development/FindAndCompareAllReleasedTransferFiles.java?revision=20387&view=markup FindAndCompareAllReleasedTransferFiles]		It is done by running the class [http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/lttoolbox-java/src/org/apertium/transfer/development/FindAndCompareAllReleasedTransferFiles.java?revision=20387&view=markup FindAndCompareAllReleasedTransferFiles]

@@ Line 1: / Line 1: @@
-This is a test of all .t1x files in all language pairs in http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/.
+This is a test of all .t1x files in all language pairs in http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/. There is 63 of them, and '''in all cases the bytecode transfer gives exactly the same result''' as the old interpreted transfer, down to the handling of wrong clip statements (but excluding invoking macros with wrong numer of parameters or parameters, as the results of that are unpredictable).
 It is done by running the class [http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/lttoolbox-java/src/org/apertium/transfer/development/FindAndCompareAllReleasedTransferFiles.java?revision=20387&view=markup FindAndCompareAllReleasedTransferFiles]
@@ Line 16: / Line 16: @@
 After compilation a corpus of 20000 sentences is runned thru. This is just an English corpus from eo-en, but it is with the English-Esperanto bidix, so the measurements should be more or less correct. Please note that if you use a symbol set very different from the English-Esperanto one, the timing results might be worth checking with your own  language data.
-Note that apertium-es-ast.es-ast.t1x doesent compile, as the heavy use of if-statements and <clip> expressions makes the bytecode becomes larger than 64kb in some rules. This can be remedied by '''caching the values of clip expressions''' in local variable, by setting the variable ''cacheClipExcpression'' to true in line 102 in http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/lttoolbox-java/src/org/apertium/transfer/compile/ParseTransferFile.java?revision=20380&view=markup , also giving a huge performance boost. Note that this should not be done for apertium-en-gl/apertium-en-gl.gl-en.t1x and apertium-en-ca/apertium-en-ca.en-ca.t1x, as they depend on the construct shown above, in line 96.
+=== es-ast and caching the values of clip expressions ===
+Note that apertium-es-ast.es-ast.t1x doesent compile, as the heavy use of if-statements and <clip> expressions makes the bytecode becomes larger than 64kb in some rules. This can be remedied by '''caching the values of clip expressions''' in local variable, by setting the variable ''cacheClipExcpression'' to true in line 102 in http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/lttoolbox-java/src/org/apertium/transfer/compile/ParseTransferFile.java?revision=20380&view=markup , also giving a considerable performance boost. Note that this should not be done for apertium-en-gl/apertium-en-gl.gl-en.t1x and apertium-en-ca/apertium-en-ca.en-ca.t1x, as they depend on the construct shown above, in line 96.

← Older revision	Revision as of 10:24, 11 October 2010
(No difference)

← Older revision		Revision as of 02:12, 28 February 2010
Line 16:		Line 16:
	After compilation a corpus of 20000 sentences is runned thru. This is just an English corpus from eo-en, but it is with the English-Esperanto bidix, so the measurements should be more or less correct. Please note that if you use a symbol set very different from the English-Esperanto one, the timing results might be worth checking with your own language data.		After compilation a corpus of 20000 sentences is runned thru. This is just an English corpus from eo-en, but it is with the English-Esperanto bidix, so the measurements should be more or less correct. Please note that if you use a symbol set very different from the English-Esperanto one, the timing results might be worth checking with your own language data.

			Note that apertium-es-ast.es-ast.t1x doesent compile, as the heavy use of if-statements and <clip> expressions makes the bytecode becomes larger than 64kb in some rules. This can be remedied by '''caching the values of clip expressions''' in local variable, by setting the variable ''cacheClipExcpression'' to true in line 102 in http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/lttoolbox-java/src/org/apertium/transfer/compile/ParseTransferFile.java?revision=20380&view=markup , also giving a huge performance boost. Note that this should not be done for apertium-en-gl/apertium-en-gl.gl-en.t1x and apertium-en-ca/apertium-en-ca.en-ca.t1x, as they depend on the construct shown above, in line 96.

@@ Line 1: / Line 1: @@
 This is a test of all .t1x files in all language pairs in http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/.
@@ Line 11: / Line 10: @@
 * Invoking macros with wrong numer of parameters
 * Invoking macros with an out-of-range parameter (e.g. <call-macro> <with-param pos="5"/> when there are only 4 words)
+Read more in [[Validating transfer files]].
 Please find your favorite language pair below and check.

@@ Line 1: / Line 1: @@
 This is a test of all .t1x files in all language pairs in http://apertium.svn.sourceforge.net/viewvc/apertium/trunk/.
@@ Line 768: / Line 767: @@
 bytecode compiled transfer took 18.97 secs
 Speedup factor: 5.11
-transferFiles = new String[] {
-};
 OK: Output of interpreted and bytecode compiled transfer is exactly the same
 Transfer files doesent validate 6: [apertium-nn-nb/dev/apertium-nn-nb.multiwords.t1x, apertium-oc-es/apertium-oc-es.es-oc.t1x, apertium-oc-es/apertium-oc-es.oc-es.t1x, apertium-br-fr/dev/old_rules.t1x, apertium-oc-ca/apertium-oc-ca.ca-oc.t1x, apertium-oc-ca/apertium-oc-ca.oc-ca.t1x]
 Interpreted (old) tranfer fails 0: []