Used cppclean to help identify useless includes and removed them. This
involved erroneously included headers, but also cases where forward
declarations could have been used rather than a full include.
the GPUCoalescer code is used in the ruby profiler regardless of
whether or not the coalescer code has been compiled, which can
lead to link/run time errors. here we add #ifdefs to guard the
usage of GPUCoalescer code. eventually we should refactor this
code to use probe points.
Garnet's NetworkInterface does not consider the size of MessageBuffers when
ejecting a Message from the network. Add a size check for the MessageBuffer
and only enqueue if space is available. If space is not available, the
message if placed in a queue and the credit is held. A callback from the
MessageBuffer is implemented to wake the NetworkInterface. If there are
messages in the stalled queue, they are processed first, in a FIFO manner
and if succesfully ejected, the credit is finally sent back upstream. The
maximum size of the stall queue is equal to the number of valid VNETs
with MessageBuffers attached.
This patch is an updated version of /r/3297.
"The most important statistic for measuring memory hierarchy performance is
throughput, which is affected by independent variables, buffer sizing and
communication latency. It is difficult/impossible to debug performance issues
through series buffers without knowing which are the bottlenecks. For finite
buffers, this patch adds statistics for the average number of messages in the
buffer, the occupancy of the buffer slots, and number of message stalls."
The NetworkInterface wakeup currently iterates over all VNETs and breaks the
loop if a VNET is unable to allocate a VC. This can cause a deadlock if a
lower numbered VNET is unable to allocate a VC while a higher numbered VNET
has idle VCs. This seems like a bug as Garnet 1.0 uses a while loop over an
if-statement, suggesting the break was intended for this while loop. This
patch removes the break statement, which allows up to one message to be
dequeued from a VNET and injected into the network.
When Ruby controllers stall messages in MessageBuffers, the buffer moves those
messages off the priority heap and into a per-address stall map. When buffers
are finite-sized, the test areNSlotsAvailable() only checks the size of the
priority heap, but ignores the stall map, so the map is allowed to grow
unbounded if the controller stalls numerous messages. This patch fixes the
problem by tracking the stall map size and testing the total number of messages
in the buffer appropriately.
the iterator declared in DMASequencer::ackCallback() is only used in an
assert, this causes clang to fail when building fast. here we move
the find call on the request table directly into the assert.
The Python wrappers generally assume that destructors are public. Make
the BaseXBar destructor public to avoid confusing the Python wrapper.
Change-Id: If958802409c0be74e875dd6e279742abfdb3ede1
Signed-off-by: Andreas Sandberg <andreas.sandberg@arm.com>
Reviewed-by: Nikos Nikoleris <nikos.nikoleris@arm.com>
Reviewed-by: Curtis Dunham <curtis.dunham@arm.com>
This patch detects garnet network deadlock by monitoring
network interfaces. If a network interface continuously
fails to allocate virtual channels for a message, a
possible deadlock is detected.
Previously when an InvalidateReq snooped a cache with a dirty block or
a pending modified MSHR, it would invalidate the block or set the
postInv flag. The cache would not send an InvalidateResp. though,
causing memory order violations. This patches changes this behavior,
making the cache with the dirty block or pending modified MSHR the
ordering point.
Change-Id: Ib4c31012f4f6693ffb137cd77258b160fbc239ca
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
Previously an MSHR with one or more invalidating targets would first
service all targets in the MSHR TargetList and then invalidate the
block. As a result any service snooping targets would lookup in the
cache and incorrectly find the block. This patch forces the
invalidation to happen when the first invalidating target is
encountered.
Change-Id: I9df15de24e1d351cd96f5a2c424d9a03d81c2cce
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
This patch changes an assertion that previously assumed that a non
invalidating snoop request should never be serviced by an
InvalidateReq MSHR. The MSHR serves as the ordering point for the
snooping packet. When the InvalidateResp reaches the cache the
snooping packet snoops the caches above to find the requested
block. One or more of the caches above will have the block since
earlier it has seen a WriteLineReq.
Change-Id: I0c147c8b5d5019e18bd34adf9af0fccfe431ae07
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
When the snoopFilter receives a response, it updates its state using
the hasSharers flag (indicates whether there are more than one copies
of the block in the caches above). The hasSharers flag of the packet
was previously populated when the request was traversing and snooping
the caches looking for the block.
1) When the response is coming from the memory-side port, its order
with respect to other responses is not necessarily preserved (e.g., a
request that arrived second to the xbar can get its response first). As
a result the snoopFilter might process responses out of order updating
its residency information using the non valid hasSharers flag which was
populated much earlier.
2) When the response is from an on-chip, the MSHRs preserve a well
defined order and the hasSharers flag should contain valid
information.
This patch changes the snoopFilter by avoiding the hasSharers flag
when the response is from the memory-side port.
Change-Id: Ib2d22a5b7bf3eccac64445127d2ea20ee74bb25b
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
Reviewed-by: Stephan Diestelhorst <stephan.diestelhorst@arm.com>
Previously, a WriteLineReq that missed in a cache would send out an
InvalidateReq if the block lookup failed or an UpgradeReq if the
block lookup succeeded but the block had sharers. This changes ensures
that a WriteLineReq always sends an InvalidateReq to invalidate all
copies of the block and satisfy the WriteLineReq.
Change-Id: I207ff5b267663abf02bc0b08aeadde69ad81be61
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
Change-Id: Ie3beeef25331f84a0a5bcc17f7a791f4a829695b
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
Reviewed-by: Stephan Diestelhorst <stephan.diestelhorst@arm.com>
This patch fixes an issue where an MSHR would incorrectly be perceived
to provide data to targets arriving after an InvalidateReq. To address
this the InvalidateReq is now treated as isForward, much like an
UpgradeReq that did not hit in the cache.
Change-Id: Ia878444d949539b5c33fd19f3e12b0b8a872275e
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
Reviewed-by: Stephan Diestelhorst <stephan.diestelhorst@arm.com>
Previously DPRINTFs printing information about a packet would use ad hoc
formats. This patch changes all DPRINTFs to use the print function
defined by the packet class, making the packet printing format more
uniform and easier to change.
Change-Id: Idd436a9758d4bf70c86a574d524648b2a2580970
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
Reviewed-by: Stephan Diestelhorst <stephan.diestelhorst@arm.com>
A response to a ReadReq can either be a ReadResp or a
ReadRespWithInvalidate. As we add targets to an MSHR for a ReadReq we
assume that the response will be a ReadResp. When the response is
invalidating (ReadRespWithInvalidate) servicing more than one targets
can potentially violate the memory ordering. This change fixes the way
we handle a ReadRespWithInvalidate. When a cache receives a
ReadRespWithInvalidate we service only the first FromCPU target and
all the FromSnoop targets from the MSHR target list. The rest of the
FromCPU targets are deferred and serviced by a new request.
Change-Id: I75c30c268851987ee5f8644acb46f440b4eeeec2
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
Reviewed-by: Stephan Diestelhorst <stephan.diestelhorst@arm.com>
Previously the information of whether a response was allocating or not
was a property of the MSHR. This change makes this flag a property of
the TargetList. Differernt TargetLists, e.g. the targets and the
deferred targets lists might have different values. Additionally, the
information about whether each of the target expects an allocating
response is stored inside the TargetList container. This allows for
repopulating the flag in case some of the targets are removed.
Change-Id: If3ec2516992f42a6d9da907009ffe3ab8d0d2021
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
Reviewed-by: Stephan Diestelhorst <stephan.diestelhorst@arm.com>
This patch adds support for repopulating the flags of an MSHR
TargetList. The added functionality makes it possible to remove
targets from a TargetList without leaving it in an inconsistent state.
Change-Id: I3f7a8e97bfd3e2e49bebad056d11bbfb087aad91
Reviewed-by: Andreas Hansson <andreas.hansson@arm.com>
Reviewed-by: Stephan Diestelhorst <stephan.diestelhorst@arm.com>
In MessageBuffer the m_not_avail_count member is incremented but not used.
This causes an overflow reported by ASAN. This patch changes from an int to
Stats::Scalar, since the count is useful in debugging finite MessageBuffers.
If the cache access mode is parallel, i.e. "sequential_access" parameter
is set to "False", tags and data are accessed in parallel. Therefore,
the hit_latency is the maximum latency between tag_latency and
data_latency. On the other hand, if the cache access mode is
sequential, i.e. "sequential_access" parameter is set to "True",
tags and data are accessed sequentially. Therefore, the hit_latency
is the sum of tag_latency plus data_latency.
Signed-off-by: Jason Lowe-Power <jason@lowepower.com>
SequencerMsg is autogenerated by slicc scripts and the MessageSizeType is
initialized to the max enume value by default. The DMASequencer pushes this
message to the mandatory queue and since the MessageSizeType is unitialized,
string_to_MessageSizeType() function used by traces to print the message fails
with a panic. This patch avoids this problem by initializing MessageSizeType
of SequencerMsg to Request_Control.
fixes to appease clang++. tested on:
Ubuntu clang version 3.5.0-4ubuntu2~trusty2
(tags/RELEASE_350/final) (based on LLVM 3.5.0)
Ubuntu clang version 3.6.0-2ubuntu1~trusty1
(tags/RELEASE_360/final) (based on LLVM 3.6.0)
the fixes address the following five issues:
1) the exec continuations in gpu_static_inst.hh were marked
as protected when they should be public. here we mark
them as public
2) the Abs instruction uses std::abs() in its execute method.
because Abs is templated, it can also operate on U32 and U64,
types, which cause Abs::execute() to pass uint32_t and uint64_t
types to std::abs() respectively. this triggers a warning
because std::abs() has no effect in this case. to rememdy this
we add template specialization for the execute() method of Abs
when its template paramter is U32 or U64.
3) Some potocols that utilize the code in cprintf.hh were missing
includes to BoolVec.hh, which defines operator<< for the BoolVec
type. This would cause issues when the generated code would try
to pass a BoolVec type to a method in cprintf.hh that used
operator<< on an instance of a BoolVec.
4) Surprise, clang doesn't like it when you clobber all the bits
in a newly allocated object. I.e., this code:
tlb = new GpuTlbEntry\[size\];
std::memset(tlb, 0, sizeof(GpuTlbEntry) \* size);
Let's use std::vector to track the TLB entries in the GpuTlb now...
5) There were a few variables used only in DPRINTFs, so we mark them
with M5_VAR_USED.
DMA sequencers and protocols can currently only issue one DMA access at
a time. This patch implements the necessary functionality to support
multiple outstanding DMA requests in Ruby.
the RequestDesc was previously implemented as a std::pair, which made
the implementation overly complex and error prone. here we encapsulate the
packet, primary, and secondary types all in a single data structure with
all members properly intialized in a ctor
Added power-down state transitions to the DRAM controller model.
Added per rank parameter, outstandingEvents, which tracks the number
of outstanding command events and is used to determine when the
controller should transition to a low power state.
The controller will only transition when there are no outstanding events
scheduled and the number of command entries for the given rank is 0.
The outstandingEvents parameter is incremented for every RD/WR burst,
PRE, and REF event scheduled. ACT is implicitly covered by RD/WR
since burst will always issue and complete after a required ACT.
The parameter is decremented when the event is serviced (completed).
The controller will automatically transition to ACT power down,
PRE power down, or SREF.
Transition to ACT power down state scheduled from:
1) The RespondEvent, where read data is received from the memory.
ACT power-down entry will be scheduled when one or more banks is
open, all commands for the rank have completed (no more commands
scheduled), and there are no commands in queue for the rank
Transition to PRE power down scheduled from:
1) respondEvent, when all banks are closed, all commands have
completed, and there are no commands in queue for the rank
2) prechargeEvent when all banks are closed, all commands have
completed, and there are no commands in queue for the rank
3) refreshEvent, after the refresh is complete when the previous
state was ACT power-down
4) refreshEvent, after the refresh is complete when the previous
state was PRE power-down and there are commands in the queue.
Transition to SREF will be scheduled from:
1) refreshEvent, after the refresh is completes when the previous
state was PRE power-down with no commands in queue
Power-down exit commands are scheduled from:
1) The refreshEvent, prior to issuing a refresh
2) doDRAMAccess, to wake-up the rank for RD/WR command issue.
Self-refresh exit commands are scheduled from:
1) The next request event, when the queue has commands for the rank
in the readQueue or there are commands for the rank in the
writeQueue and the bus state is WRITE.
Change-Id: I6103f660776e36c686655e71d92ec7b5b752050a
Reviewed-by: Radhika Jagtap <radhika.jagtap@arm.com>
The per rank statistics are periodically updated based on
state transition and refresh events.
Add a method to update these when a dump event occurs to
ensure they reflect accurate values.
Specifically, need to ensure that the low-power state
durations, power, and energy are logged correctly.
Change-Id: Ib642a6668340de8f494a608bb34982e58ba7f1eb
Reviewed-by: Radhika Jagtap <radhika.jagtap@arm.com>
Add constraint that all ranks have to be in PWR_IDLE
before signaling drain complete
This will ensure that the banks are all closed and the rank
has exited any low-power states.
On suspend, update the power stats to sync the DRAM power logic
The logic maintains the location of the signalDrainDone
method, which is still triggered from either:
1) Read response event
2) Next request event
This ensures that the drain will complete in the READ bus
state and minimizes the changes required.
Change-Id: If1476e631ea7d5999fe50a0c9379c5967a90e3d1
Reviewed-by: Radhika Jagtap <radhika.jagtap@arm.com>
Add local variable to stores commands to be issued.
These commands are in order within a single bank but will be out
of order across banks & ranks.
A new procedure, flushCmdList, sorts commands across banks / ranks,
and flushes the sorted list, up to curTick() to DRAMPower.
This is currently called in refresh, once all previous commands are
guaranteed to have completed. Could be called in other events like
the powerEvent as well.
By only flushing commands up to curTick(), will not get out of sync
when flushed at a periodic stats dump (done in subsequent patch).
Change-Id: I4ac65a52407f64270db1e16a1fb04cfe7f638851
Reviewed-by: Radhika Jagtap <radhika.jagtap@arm.com>
This patch adds port direction names to the links during topology
creation, which can be used for better printed names for the links
or for users to code up their own adaptive routing algorithms.
It also adds support for every router to have an independent latency
value to support heterogeneous topologies with the subsequent
garnet2.0 patch.
This patch makes the internal links within the network topology
unidirectional, thus allowing any deadlock-free routing algorithms to
be specified from the topology itself using weights.
This patch also renames Mesh.py and MeshDirCorners.py to
Mesh_XY.py and MeshDirCorners_XY.py (Mesh with XY routing).
It also adds a Mesh_westfirst.py and CrossbarGarnet.py topologies.
Sanchayan Maity
