<?xml version="1.0" encoding="UTF-8"?>
<rss xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#" xmlns:taxo="http://purl.org/rss/1.0/modules/taxonomy/" version="2.0">
  <channel>
    <title>topic Performance comparison for Convolution - Direct vs FFT vs Auto in Intel® oneAPI Math Kernel Library</title>
    <link>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774711#M906</link>
    <description>Hi,&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;This is regarding the comparison of Direct Mode vs FFT mode for Multi-dimensional convolutions in terms of performance. My expectation was that the FFT mode should be faster than direct mode, but it turned out to be the other way for my test cases.&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;I tested all modes for 2-dimensional convolution of a 256 x 256 matrix with a 5 x 5 kernel&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;Approximate runtime (averaged over 1000 runs) was&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;DIRECT mode: 2.4 ms&lt;/DIV&gt;&lt;DIV&gt;FFT mode: 9.5 ms&lt;/DIV&gt;&lt;DIV&gt;AUTO mode: 9.5 ms&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;I was surprised by the slower run time of FFT mode.Any insight regarding performance of Intel MKL convolution routines will be helpful.&lt;/DIV&gt;&lt;DIV&gt;&lt;BR /&gt;Thanks,&lt;/DIV&gt;&lt;DIV&gt;&lt;BR /&gt;Rahul Modi.&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;</description>
    <pubDate>Tue, 10 Apr 2012 20:28:52 GMT</pubDate>
    <dc:creator>rahul_modi</dc:creator>
    <dc:date>2012-04-10T20:28:52Z</dc:date>
    <item>
      <title>Performance comparison for Convolution - Direct vs FFT vs Auto</title>
      <link>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774711#M906</link>
      <description>Hi,&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;This is regarding the comparison of Direct Mode vs FFT mode for Multi-dimensional convolutions in terms of performance. My expectation was that the FFT mode should be faster than direct mode, but it turned out to be the other way for my test cases.&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;I tested all modes for 2-dimensional convolution of a 256 x 256 matrix with a 5 x 5 kernel&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;Approximate runtime (averaged over 1000 runs) was&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;DIRECT mode: 2.4 ms&lt;/DIV&gt;&lt;DIV&gt;FFT mode: 9.5 ms&lt;/DIV&gt;&lt;DIV&gt;AUTO mode: 9.5 ms&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;I was surprised by the slower run time of FFT mode.Any insight regarding performance of Intel MKL convolution routines will be helpful.&lt;/DIV&gt;&lt;DIV&gt;&lt;BR /&gt;Thanks,&lt;/DIV&gt;&lt;DIV&gt;&lt;BR /&gt;Rahul Modi.&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;</description>
      <pubDate>Tue, 10 Apr 2012 20:28:52 GMT</pubDate>
      <guid>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774711#M906</guid>
      <dc:creator>rahul_modi</dc:creator>
      <dc:date>2012-04-10T20:28:52Z</dc:date>
    </item>
    <item>
      <title>Performance comparison for Convolution - Direct vs FFT vs Auto</title>
      <link>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774712#M907</link>
      <description>Hi Rahul,&lt;BR /&gt;&lt;BR /&gt;Which MKL release and Intel processordo you use? &lt;BR /&gt;&lt;BR /&gt;FFT mode vs. DIRECT mode:&lt;BR /&gt;FFT mode hasoverhead, soexecution time of FFT modecan be bigger then execution time of direct mode for "small" problem. This overhead depends to hardware, MKL version and sizes of problem.&lt;BR /&gt;&lt;BR /&gt;Victor.</description>
      <pubDate>Wed, 11 Apr 2012 03:24:22 GMT</pubDate>
      <guid>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774712#M907</guid>
      <dc:creator>Victor_Gladkikh</dc:creator>
      <dc:date>2012-04-11T03:24:22Z</dc:date>
    </item>
    <item>
      <title>Performance comparison for Convolution - Direct vs FFT vs Auto</title>
      <link>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774713#M908</link>
      <description>Hey Victor,&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;Thanks again for your prompt reply.&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;I am using MKL ver 10.3 Update 8 release and I am using an Intel core i7-2600K CPU @ 3.4 GHz&lt;/DIV&gt;&lt;DIV&gt;Currently I have MKL configured to run in the sequential mode.&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;So do you suggest using direct mode for smaller problems?&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;As for more on the test problem, I had a simple 256*256 input array&lt;/DIV&gt;&lt;DIV&gt;0&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;1&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;2&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;. . . .&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;255&lt;/DIV&gt;&lt;DIV&gt;256 257 ... ... ... ...  511&lt;/DIV&gt;&lt;DIV&gt;...&lt;/DIV&gt;&lt;DIV&gt;...&lt;/DIV&gt;&lt;DIV&gt;65280 ... ... ... ...   65535 &lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;Ans this array is convolved with a 5*5 kernel&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;&lt;DIV id="_mcePaste"&gt;-1,&lt;SPAN style="white-space: pre;"&gt;		&lt;/SPAN&gt;-0.5,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0.5,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;1,&lt;/DIV&gt;&lt;DIV id="_mcePaste"&gt;					 -1.5,&lt;SPAN style="white-space: pre;"&gt;		&lt;/SPAN&gt;-0.75,
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0.75,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;1.5,&lt;/DIV&gt;&lt;DIV id="_mcePaste"&gt;					 -2,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;-1,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;1,&lt;SPAN style="white-space: pre;"&gt;		&lt;/SPAN&gt;2,&lt;/DIV&gt;&lt;DIV id="_mcePaste"&gt;					 -1.5,
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;-0.75,
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0.75,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;1.5,&lt;/DIV&gt;&lt;DIV id="_mcePaste"&gt;					 -1,
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;-0.5,
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0.5,&lt;SPAN style="white-space: pre;"&gt;		&lt;/SPAN&gt;1,&lt;/DIV&gt;-1,
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;-0.5,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;0.5,&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;
&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;1&lt;SPAN style="white-space: pre;"&gt;	&lt;/SPAN&gt;&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;Thanks for all your help.&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;&lt;DIV&gt;Best,&lt;/DIV&gt;&lt;DIV&gt;Rahul Modi.&lt;/DIV&gt;&lt;DIV&gt;&lt;/DIV&gt;</description>
      <pubDate>Wed, 11 Apr 2012 20:41:55 GMT</pubDate>
      <guid>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774713#M908</guid>
      <dc:creator>rahul_modi</dc:creator>
      <dc:date>2012-04-11T20:41:55Z</dc:date>
    </item>
    <item>
      <title>Performance comparison for Convolution - Direct vs FFT vs Auto</title>
      <link>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774714#M909</link>
      <description>&lt;P&gt;Teoretically direct mode is faster then FFT mode for kernels less then 7*7but actuallycut-off pointdepends to performance of FFT mode and direct mode which depend to MKL version andhardaware. &lt;BR /&gt;&lt;BR /&gt;Victor&lt;/P&gt;</description>
      <pubDate>Thu, 12 Apr 2012 03:53:26 GMT</pubDate>
      <guid>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774714#M909</guid>
      <dc:creator>Victor_Gladkikh</dc:creator>
      <dc:date>2012-04-12T03:53:26Z</dc:date>
    </item>
    <item>
      <title>Performance comparison for Convolution - Direct vs FFT vs Auto</title>
      <link>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774715#M910</link>
      <description>Thanks, that helps&lt;DIV&gt;-Rahul&lt;/DIV&gt;</description>
      <pubDate>Thu, 12 Apr 2012 12:51:42 GMT</pubDate>
      <guid>https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Performance-comparison-for-Convolution-Direct-vs-FFT-vs-Auto/m-p/774715#M910</guid>
      <dc:creator>rahul_modi</dc:creator>
      <dc:date>2012-04-12T12:51:42Z</dc:date>
    </item>
  </channel>
</rss>

