Hi Tsung, That helper function is actually used by all vector instructions that have SDWA support, so posting the error function does not tell us which specific instruction is causing the failure. We would need a gdb backtrace for that. However, I don't think getting a gdb backtrace is the most important thing at the moment. Instead, can you please tell us more about how you determined that the original failure was caused by this SDWA failure? If the SDWA failure was happening all along, then shouldn't you have seen the assert failure in the SDWA helper function instead of the original error you posted?
Thanks, Matt <http://cs.wisc.edu/~sinclair> On Mon, Aug 13, 2018 at 6:32 PM, Tsungtai Yeh <[email protected]> wrote: > Matt, > > > 1. This error occurs in DeepBench RNN LSTM app. > 2. The (hidden size, batch size, time step) is (256, 1, 1), (512, 1, > 1), and (1024, 1, 1) > 3. This error happens in MIOpenGEMM "miog_alphaab" kernel > 4. I don't think this error caused by atomic operation, but I can check. > > > Tsung Tai > > ________________________________ > From: Tsungtai Yeh > Sent: Monday, August 13, 2018 2:20:44 PM > To: [email protected] > Cc: [email protected] > Subject: [gem5-dev] RubyPort::MemSlavePort::recvTimingReq(PacketPtr): > Assertion `getOffset(pkt->getAddr()) + pkt->getSize() <= > RubySystem::getBlockSizeBytes()' failed on X86_GCN3. > > > The following debug trace tells us the "getOffset(pkt->getAddr()) + > pkt->getSize() <= RubySystem::getBlockSizeBytes()". Why does this > exception happen? Do you give me some hints I can continue to dig in this > error? Thank you. > > > 1428070516000: system.ruby.tcp_cntrl2.coalescer.slave0: > MemSlavePort::recvTimingReq: pkt->getAddr:#20f5c3d, getOffset:#61, > BlockSize:64, pkt->getSize:4, print:ReadReq [20f5c3d:20f5c40] > > > gem5.opt: build/GCN3_X86/mem/ruby/system/RubyPort.cc:294: virtual bool > RubyPort::MemSlavePort::recvTimingReq(PacketPtr): Assertion > `getOffset(pkt->getAddr()) + pkt->getSize() <= > RubySystem::getBlockSizeBytes()' failed. > _______________________________________________ > gem5-dev mailing list > [email protected] > http://m5sim.org/mailman/listinfo/gem5-dev > _______________________________________________ gem5-dev mailing list [email protected] http://m5sim.org/mailman/listinfo/gem5-dev
