You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@systemml.apache.org by "Niketan Pansare (JIRA)" <ji...@apache.org> on 2017/10/26 20:42:00 UTC
[jira] [Commented] (SYSTEMML-1975) Fix GPU tests for single
precision
[ https://issues.apache.org/jira/browse/SYSTEMML-1975?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16221207#comment-16221207 ]
Niketan Pansare commented on SYSTEMML-1975:
-------------------------------------------
Following tests are failing for double precision:
{code}
Failed tests:
MatrixMultiplicationOpTest.matrixMatrixTest2:161->assertMatrixMultiplication:261->GPUTests.assertEqualObjects:367->GPUTests.assertEqualMatrices:267 Relative error(1.000000) is more than threshold (0.000010). Expected = -0.002456, Actual = 0.000000, differed at [0, 0]
MatrixMultiplicationOpTest.transposeSelfMatrixMultiply:235->GPUTests.assertEqualObjects:367->GPUTests.assertEqualMatrices:272 expected:<0.0> but was:<1.929108490094312>
{code}
My guess is there is likely a bug caught by these two tests and other tests are likely related to precision.
> Fix GPU tests for single precision
> ----------------------------------
>
> Key: SYSTEMML-1975
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1975
> Project: SystemML
> Issue Type: Bug
> Reporter: Niketan Pansare
>
> The current GPU tests for single precision compares the results of double precision operation on CPU with single precision operation on GPU. So, it is difficult to compare the results. When we support single precision on CPU, we can rework the GPU tests to fix the following tests:
> {code}
> Failed tests:
> NeuralNetworkOpTests.testConv2dBackwardFilter:324->GPUTests.assertEqualObjects:367->GPUTests.assertEqualMatrices:269 Relative error(0.000012) is more than threshold (0.000010). Expected = 25311.000000, Actual = 25310.687500, differed at [0, 50]
> MatrixMultiplicationOpTest.matrixMatrixTest2:161->assertMatrixMultiplication:261->GPUTests.assertEqualObjects:367->GPUTests.assertEqualMatrices:269 Relative error(1.000000) is more than threshold (0.000010). Expected = -0.002456, Actual = 0.000000, differed at [0, 0]
> MatrixMultiplicationOpTest.matrixMatrixTest4:209->assertMatrixMultiplication:261->GPUTests.assertEqualObjects:367->GPUTests.assertEqualMatrices:269 Relative error(0.000024) is more than threshold (0.000010). Expected = 0.573809, Actual = 0.573822, differed at [5, 127]
> MatrixMultiplicationOpTest.commonCaseDLMatrixMatrixTest2:137->assertMatrixMultiplication:261->GPUTests.assertEqualObjects:367->GPUTests.assertEqualMatrices:269 Relative error(0.000013) is more than threshold (0.000010). Expected = -0.922950, Actual = -0.922939, differed at [0, 13]
> MatrixMultiplicationOpTest.transposeSelfMatrixMultiply:235->GPUTests.assertEqualObjects:367->GPUTests.assertEqualMatrices:272 expected:<0.0> but was:<1.9291085004806519>
> ScalarMatrixElementwiseOpTests.testModulusLeftScalar:167->runScalarMatrixElementWiseTests:261->GPUTests.assertEqualObjects:367->GPUTests.assertEqualMatrices:269 Relative error(1.000000) is more than threshold (0.000010). Expected = 0.078125, Actual = 0.000000, differed at [1500, 22]
> ScalarMatrixElementwiseOpTests.testIntegerDivideLeftScalar:157->runScalarMatrixElementWiseTests:261->GPUTests.assertEqualObjects:367->GPUTests.assertEqualMatrices:269 Relative error(0.015873) is more than threshold (0.000010). Expected = 63.000000, Actual = 64.000000, differed at [1500, 22]
> {code}
> [~nakul02]
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)