gpu-compute: No RF scheduling in case of SKIP or EMPTY
authorAlexandru Dutu <alexandru.dutu@amd.com>
Wed, 11 Jul 2018 20:32:18 +0000 (16:32 -0400)
committerAnthony Gutierrez <anthony.gutierrez@amd.com>
Fri, 17 Jul 2020 16:34:59 +0000 (16:34 +0000)
commit7d50d5d97237c2df0c08ccfbe6e5c40591d3bcff
tree9cb59f88a3a3f56400111f7606915971137d7afc
parent5f0378b8d00ef7b133e0232fad409b2b65093a8b
gpu-compute: No RF scheduling in case of SKIP or EMPTY

In case of flat memory instructions the status for the
LM pipe execution unit is set to SKIP or EMPTY, as the bus
between the VRF and the GM and LM pipe is shared. The
destination operands should not be scheduled for the LM pipe,
event if the wave is in the dispatch list. This can lead
to deadlock in the destination cache as DCEs are reused
and the slotsAvailableForBank count gets artificially
incremented.

Change-Id: I2230c53e3bc1032d2cccbe00fab62c99ab8de6cd
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/29970
Maintainer: Anthony Gutierrez <anthony.gutierrez@amd.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
src/gpu-compute/schedule_stage.cc