The ROM field was originally intended as a future alternate way to load
VBIOS without the ROM being on the disk image. This code path is never
taken for the devices gem5 supports and there is no gem5 implementation.
Deprecate the rom_binary field for this reason.
Similarly, MMIO traces were only used for Vega10. Deprecate this as
Vega10 is now deprecated. The MMIO trace reader is kept as it may still
be useful in the future. It is still the primary way to handle devies
which have graphics capability. None of the devices supported by gem5
have graphics now that Vega10 is deprecated.
Invalidate requests align to system cache line size. This causes
problems if the GPU cache hierarchy's cache line size is different than
the system as the unlaigned requests never return, leading to deadlock
on deferred dispatch.
This commit uses the cache line size from the GPU memory manager and
makes the cache line size there non-optional.
Tested with multiple RubySystems where CPU side was 64B and GPU side was
128B cache lines.
Unmap queues with queue_sel of 2 unmaps all queues while queue_sel of 3
unmaps all non-static queues. The implementation of 3 was actually
correct for 2. Static queues are queues which were mapped using a map
queues packet with a queue_type of 1 or 2.
This commit adds ability to mark a queue as static. When unmap queues
with queue_sel of 2 is sent, the existing code is now executed. With a
value of 3, we now check if the queue was marked static and do not
unmap it if marked.
Change-Id: I87d7cf78a0600c7baa516c01f42c294d3c4e90c5
SDMA packets which use dmaVirtWrites call their completion event before
the write takes place in the Ruby protocol. This causes a use-after-free
issue corruption random memory locations leading to random errors. This
commit adds a cleanup event for each packet that uses DMA and sets the
cleanup latency as 10000 ticks. In atomic mode, the writes complete
exactly 2000 ticks after the completion event is called and therefore a
fixed latency can be used. This is not tested with timing mode, which
does not work with GPUFS at the moment, so a warning is added to give an
idea where to look in case the same issue occurs once timing mode is
supported.
Change-Id: I9ee2689f2becc46bb7794b18b31205f1606109d8
The SDMA engine copies data in chunks. It currently uses the pointer
returned from new[] and manipulates it using pointer arithmetic. This
modified pointer is then passed to the completion function which deletes
the pointer. Since it is not the original pointer allocated by new[]
this triggers issues in ASAN.
Change-Id: I03ccf026633285e75005509445c62fcbda8eb978
SDMA RLC queues do not currently remove their doorbell mapping. This can
cause issues re-registering the queue and prevents the pending doorbells
feature from working. In addition the data value of the doorbell (the
ring buffer rptr) is not saved, leading to UB when this workaround is
used.
This commit removes the doorbell mapping from the gpu device when the
SDMA engine unmaps an RLC queue and copies the next doorbell value to
the pending packet as was originally intended.
Change-Id: Ifd551450f439c065579afcf916f8ff192e7598ab
This is the version for MI300. For the most part, it is the same as
MI200 with the exception of architected flat scratch (not yet
implemented in gem5) and therefore a new version enum is required.
Change-Id: Id18cd7b57c4eebd467c010a3f61e3117beb8d58a
SDMA ptePde packets are generating a warning that a GART address is
missing, causing a wrong address to be clobbered by the operation.
This commit fixes this by converting the GART address when the queue is
running in privledged mode, which is the only mode allowed to use GART
addresses. This removes the warnings and writes to the correct memory
region.
Change-Id: I64acac308db2431c5996b876bf4cda704f51cf25
Fixes several memory leaks, mostly of small and medium severity. Fixes
mismatched new/new[] and delete/delete[] calls.
Change-Id: Iedafc409389bd94e45f330bc587d6d72d1971219
The GPU device currently supports large BAR which means that the driver
can write directly to GPU memory over the PCI bus without using SDMA or
PM4 packets. The gem5 PCI interface only provides an atomic interface
for BAR reads/writes, which means the values cannot go through timing
mode Ruby caches. This causes bugs as the TCC cache is allowed to keep
clean data between kernels for performance reasons. If there is a BAR
write directly to memory bypassing the cache, the value in the cache is
stale and must be invalidated.
In this commit a TCC invalidate is generated for all writes over PCI
that go directly to GPU memory. This will also invalidate TCP along the
way if necessary. This currently relies on the driver synchonization
which only allows BAR writes in between kernels. Therefore, the cache
should only be in I or V state.
To handle a race condition between invalidates and launching the next
kernel, the invalidates return a response and the GPU command processor
will wait for all TCC invalidates to be complete before launching the
next kernel.
This fixes issues with stale data in nanoGPT and possibly PENNANT.
Change-Id: I8e1290f842122682c271e5508a48037055bfbcdf
Currently gem5 assumes that there is only one command processor (CP)
which contains the PM4 packet processor. Some GPU devices have multiple
CPs which the driver tests individually during POST if they are used or
not. Therefore, these additional CPs need to be supported.
This commit allows for multiple PM4 packet processors which represent
multiple CPs. Each of these processors will have its own independent
MMIO address range. To more easily support ranges, the MMIO addresses
now use AddrRange to index a PM4 packet processor instead of the
hard-coded constexpr MMIO start and size pairs.
By default only one PM4 packet processor is created, meaning the
functionality of the simulation is unchanged for devices currently
supported in gem5.
Change-Id: I977f4fd3a169ef4a78671a4fb58c8ea0e19bf52c
PCIe can read/write to any 32-bit address using the PCI index/index2
registers as an address and then reading/writing the corresponding
data/data2 register.
This commit adds this functionality and removes one magic value being
written to support GPU POST. This feature is disabled for Vega10 which
relies on an MMIO trace for too many values to implement in the MMIO
interface.
Change-Id: Iacfdd1294a7652fc3e60304b57df536d318c847b
The SRBM write packets where previously not required. This commit
implements SRBM writes to set a register by using the new setRegVal
interface. SRBM writes seem to be used for SRIOV enabled devices.
Change-Id: I202653d339e882e8de59d69a995f65332b2dfb8c
The top level AMDGPUDevice currently reads/writes all unknown registers
to/from a map containing the previously written value. This is intended
as a way to handle registers that are not part of the model but the
driver requires for functionality. Since this is at the top level, it
can mask changes to register values which do not go through the same
interface. For example, reading an MMIO, changing via PM4 queue, and
reading again returns the stale cached value.
This commit removes the usage of the regs map in AMDGPUDevice,
implements some important MMIOs that were previously handled by it, and
moves the unknown register handling to the NBIO aperture only. To reduce
the number of additional MMIOs to implement, the display manager in
vega10 is now disabled.
Change-Id: Iff0a599dd82d663c7e710b79c6ef6d0ad1fc44a2
The SDMA engine can potentially be used to write to the GART address
range. Since gem5 has a shadow copy of the GART table to avoid sending
functional reads to device memory, the GART table must be updated when
copying to the GART range.
This changeset adds a check in the VM for GART range and implements the
SDMA copy packet writing to the GART range. A fatal is added to write
and ptePde, which are the only other two ways to write to memory, as
using these packets to update the GART table has not been observed.
Change-Id: I1e62dfd9179cc9e987659e68414209fd77bba2bd
The write data packet can write multiple dwords but currently always
assumes there is one dword, which can cause some write data to be
missed. This case is not common, but the number of dwords is implicitly
defined in the PM4 header.
This changeset passes the PM4 header to write data so that the correct
number of dwords can be determined. For now we assume no page crossing
when writing multiple dwords as the driver should be checking for that.
Change-Id: I0e8c3cbc28873779f468c2a11fdcf177210a22b7
The ROCm 6.0 driver adds a node_id field to interrupts which must match
before passing on the interrupt to be cleared by the cookie from gem5's
interrupt handler implementation. Add this field and enable for gfx942.
The usage of the field can be seen in event_interrupt_isr_v9_4_3 at
https://github.com/ROCm/ROCK-Kernel-Driver/blob/roc-6.0.x/drivers/
gpu/drm/amd/amdkfd/kfd_int_process_v9.c#L449
Change-Id: Iae8b8f0386a5ad2852b4a3c69f2c161d965c4922
If the NOP count of an SDMA NOP packet goes beyond the wptr address, the
queue decode method will loop infinitely. If a packet comes in with a
bad count this causes gem5 to hang. This change advances the rptr one
dword at a time until either reaching the NOP count or when rptr == wptr
to prevent this issue.
Change-Id: Ib2c0f74a477bff27890c9c064bb4190e76e513bd
Related to issue #703 , this PR removes GCN3 related files and updates
source code, documentation, and tests to switch over to Vega is that was
not done already. Highlights are:
- Remove all src/arch/amdgpu/gcn3 files and update Kconfigs.
- Remove references to GCN3 and replace with Vega where applicable.
- Update the build targets in the gcn-gpu Docker. This will need to be
rebuilt but not urgently.
- Remove the GCN3 tag in testlib. Most tests seem to be using Vega
already, so that commit is small.
By default all SDMA queues are privileged queues, meaning the addresses
in SDMA packets use the privileged translation tables. RLC queues
(sometimes called user queues) are not necessarily privileged and might
use user translation tables. RLC queues are used more often in ROCm 6.0
exposing an issue with invalid translations with RLC queues.
This changeset checks the priv bit in the SDMA MQD when an RLC queue is
mapped. Each packet type which uses an address then checks the bit
before performing translation. Tested with daily/weekly tests with a
ROCm 6.0 disk image and tests are passing.
Change-Id: I6122fbc194e8d6f5d38e81f1b0e11646d90e0ea0
Replace instances of "GCN3" with Vega. Remove gfx801 and gfx803. Rename
FIJI to Vega and Carrizo to Raven.
Using misc since there is not enough room to fit all the tags.
Change-Id: Ibafc939d49a69be9068107a906e878408c7a5891
The GPU device keeps a local copy of each ring buffers read pointer
(rptr) to avoid constant DMAs to/from host memory. This means it needs
to be periodically updated on the host side as the driver uses this to
determine how much space is left in the queue and may hang if it believe
the queue is full. For user-mode queues, this already happens when
queues are unmapped. For kernel mode queues (e.g., HIQ, KIQ) the rptr is
never updated leading to a hang.
In this patch the rptr for *all* queues is reported back to the kernel
whenever the queue reaches an empty state (rptr == wptr). Additionally
to handle PM4 queue wrap-around, the queue processing function checks if
the queue is not empty instead of rptr < wptr. This is state because the
driver fills PM4 queues with NOP packets on initialization and when wrap
around occurs.
Change-Id: Ie13a4354f82999208a75bb1eaec70513039ff30f
When restoring checkpoints for certain applications, gem5 tries to
create new doorbells with a pre-existing queue ID and simulation crashes
shortly after. This commit checkpoints the existing VMID map so that any
new doorbells after restoration use a unique queue ID
Change-Id: I9bf89a2769db26ceab4441634ff2da936eea6d6f
When restoring checkpoints for certain applications, gem5 tries to
create new doorbells with a pre-existing queue ID and simulation crashes
shortly after. This commit adds existing IDs to the GPU device's used
VMID map so that new doorbells are aware of existing queue IDs and use a
new ID. This ensures that queue IDs are unique after checkpoint
restoration
Change-Id: I9bf89a2769db26ceab4441634ff2da936eea6d6f
The amdgpu driver can, at *any* time, tell the device to unmap a queue
to force the queue descriptor to be written back to main memory in the
form of a memory queue descriptor (MQD). It will then immediately remap
the queue and continue writing the doorbell to the queue. It is possible
that the doorbell write occurs after the queue is unmapped but before it
is remapped. In this situation, we need to check the updated value of
the doorbell for the queue and write that to the queue after it is
mapped.
To handle this, a pending doorbell packet map is created to hold a
packet to replay when the queue is mapped. Because PCI in gem5
implements only the atomic protocol port, we cannot use the original
packet as it must respond in the same Tick. This patch fixes issues with
the doorbell maps not being cleared on unmapping to ensure the doorbell
is not found in writeDoorbell and places in the pending doorbell map.
This includes fixing the doorbell offset value in the doorbell to VMID
map which was is now multiplied by four as it is a dword address.
This was tested using tensorflow 2.0's MNIST example which was seeing
this issue consistently. With this patch it now makes progress and does
issue pending doorbell writes.
Change-Id: Ic6b401d3fe7fc46b7bcbf19a769cdea6814e7d1e
Earlier, GPU checkpointing was working only if a checkpoint was created
before the first kernel execution. This pull request adds support to
checkpoint in-between any two kernel calls. It does so by doing the
following.
- Adds flush support in the GPU_VIPER protocol
- Adds flush support in the GPUCoalescer
- Updates cache recorder to use the GPUCoalescer during simulation
cooldown and cache warmup times.
The ROCr runtime uses a combination of HSA signal timestamps and
hardware MMIOs to calculate profiling times. At the beginning of an
application a timestamp is read from the GPU using MMIOs. The clock
MMIOs reside in the GFX MMIO region, so a new AMDGPUGfx class is added
to handle these MMIOs.
The timestamp value is expected to be in nanoseconds, so we simply use
the gem5 tick converted to ns.
Change-Id: I7d1cba40d5042a7f7a81fd4d132402dc11b71bd4
During checkpoint restoration, the unserialize() function writes rptr,
wptr, and indirect buffer rptr, wptr to PM4 queue's rptr, wptr fields.
This commit updates this to write only the relevant pointers to the
queue structure. If indirect buffers are used, then it writes only the
indirect buffer pointers to the queue. If they are not used, then it
writes rptr, wptr values to the queue.
Change-Id: Iedb25a726112e1af99cc1e7bc012de51c4ebfd45
GPUFS uses aql information from PM4 queues to initialize doorbells. This
commit adds aql information to the checkpoint so that it can be used
during restoration to correctly initialize all doorbells. Additionally,
this commit also sets the hsa queue correctly during checkpoint-restoration
Change-Id: Ief3ef6dc973f70f27255234872a12c396df05d89
It is possible to execute a GPU atomic instruction using a memory
address that is in the host memory space (e.g, HMM, __managed__,
hipHostMalloc'd address). Since these are in host memory they are passed
to the SystemHub DmaDevice. However, this currently executes as a write
packet without modifying data. This leads to hangs in applications that
use atomics for forward progress (e.g., HeteroSync).
It is not clear where these are handled on a real GPU, but they are
certianly not handled by the software stack nor driver, so they must be
handled in hardware and therefore implemented in gem5. Handling for
atomics in the SystemHub makes the most sense.
To make atomics work a few extra changes need to be made to the
SystemHub. (1) The atomic is implemented as a host memory read, followed
by calling the AtomicOpFunctor, followed by a write. This requires a
second event to handle read response, performing atomic, and issuing a
write. (2) Atomics must be serialized otherwise two atomics might return
the same value which is incorrect. This patch adds serialization logic
for all request types to the same address to handle this. (3) With the
added complexity of the SystemHub, a new debug flag explicitly for
SystemHub is added.
Testing done: The heterosync application with input "sleepMutex 10 16 4"
previously hung before this patch. It passes with the patch applied.
This application tests both (1) and (2) above, as it allocates locks
with hipHostMalloc and has multiple workgroups sending an atomic request
in the same Tick, verifying the serialization mechanism.
Change-Id: Ife84b30037d1447dd384340cfeb06fdfd472fff9
The ROCm stack requires PCI express atomics. Currently the first PCI
CapabilityPtr does not point to anything, which signals to the OS
(Linux) that this is an early generation PCI device. As PCI express
atomics were introduced later, the CapabilityPtr needs to point to at
least a PCI express capability structure. This capability is defined as
0x10 in Linux. We additionally set the PCI atomic based bits and
implement device specific PCI configuration space reads and writes to
the amdgpu device.
With this commit, the output of simulation when loading the amdgpu
driver no longer outputs "PCIE atomics not supported". Further, an
application which uses PCIe atomics (PyTorch with a reduce_sum kernel)
now makes further progress.
Change-Id: I5e3866979659a2657f558941106ef65c2f4d9988
This SDMA packet is much more common starting around ROCm 5.4.
Previously this was mostly used to clear page tables after an
application ended and was therefore left unimplemented. It is
now used for basic operation like device memsets.
This patch implements constant fill as it is now necessary.
Change-Id: I9b2cf076ec17f5ed07c20bb820e7db0c082bbfbc
When using the new operator, delete should be called
on any allocated memory after it's use is complete.
Change-Id: Id5fcfb264b6ddc252c0a9dcafc2d3b020f7b5019
The PCI read/write functions are atomic functions in gem5, meaning they
expect a response with a latency value on the same simulation Tick. For
reads to a PCI device, the response must also include a data value read
from the device.
The AMDGPU device has a PCI BAR which mirrors the frame buffer memory.
Currently reads are done atomically, but writes are sent to a DMA device
without waiting for a write completion ACK. As a result, it is possible
that writes can be queued in the DMA device long enough that another
read for a queued address arrives. This happens very deterministically
with the AtomicSimpleCPU and causes GPUFS to break with that CPU.
This change makes writes to the frame BAR atomic the same as reads. This
avoids that problem and as a result the AtomicSimpleCPU can now load the
driver for GPUFS simulations.
Change-Id: I9a8e8b172712c78b667ebcec81a0c5d0060234db
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/71898
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matthew Poremba <matthew.poremba@amd.com>
Reviewed-by: Matthew Poremba <matthew.poremba@amd.com>
Patch https://gem5-review.googlesource.com/c/public/gem5/+/70040 added
support for a variable number of SDMA engines to support newer GPU
models. As part of this an SDMA IDs map was added to map from SDMA ID
number to the SDMA SimObject pointer. In order to get the correct
pointer in unserialize now, we need to store the ID in the checkpoint
and use that to index the new map. We can't simply assign using the loop
variable as the SDMAs might not be in order in the checkpoint and
additionally the checkpoint contains both the gfx and page offset for
the SDMA engines, so each SDMA is inserted into the SDMA offset map
(sdmaEngs) twice.
Change-Id: I08e9a8d785f467b6eebff8ab0a9336851c87258d
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/70878
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Currently gem5 assumes the amdgpu device to be Vega10. In order to
support more devices we need to handle situations where different
registers and addresses have the same functionality but different
offsets on different devices.
This changeset adds an NBIO class to handle device discovery and driver
initialization related tasks, pulling them out of the AMDGPUDevice
class. The offsets used for MMIOs are reworked slightly to use offsets
rather than absolute addresses. This is because we cannot determine the
absolute address in the constructor since the BAR has not been assigned
by the OS yet.
Change-Id: I14b364374e086e185978334425a4e265cf2760d0
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/70041
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Currently the amdgpu simulated device is assumed to be a Vega10. As a
result there are a few things that are hardcoded. One of those is the
number of SDMAs. In order to add a newer device, such as MI100+, we need
to enable a flexible number of SDMAs.
In order to support a variable number of SDMAs and with the MMIO offsets
of each device being potentially different, the MMIO interface for SDMAs
is changed to use an SDMA class method dispatch table with forwards a
32-bit value from the MMIO packet to the MMIO functions in SDMA of the
format `void method(uint32_t)`. Several changes are made to enable this:
- Allow the SDMA to have a variable MMIO base and size. These are
configured in python.
- An SDMA class method dispatch table which contains the MMIO offset
relative to the SDMA's MMIO base address.
- An updated writeMMIO method to iterate over the SDMA MMIO address
ranges and call the appropriate SDMA MMIO method which matches the
MMIO offset.
- Moved all SDMA related MMIO data bit twiddling, masking, etc. into
the MMIO methods themselves instead of in the writeMMIO method in
SDMAEngine.
Change-Id: Ifce626f84d52f9e27e4438ba4e685e30dbf06dbc
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/70040
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
If an MMIO was previously written and the driver reads it, we should
return the value that was previously read. This overwrites the MMIO
trace value which is the last resort fallback for finding an MMIO value.
This is needed to initialize newer GPU devices in gem5.
Change-Id: Ida2435290b706288e88518b5d920691cdb6dcc09
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/70039
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
During GPUFS checkpoint restore, doorbells callbacks are created based
on certain MQD attributes. These callbacks are required to create new
SDMA doorbells. If these attributes are not present in the checkpoint,
the restore hangs indefinitely waiting for ioctl calls that access these
doorbells to finish execution. This commit adds the attributes required
for checkpoint restore to proceed.
Change-Id: Id3d1b7a2627d4c50133d923096495957a233f675
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/70077
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Reviewed-by: Matthew Poremba <matthew.poremba@amd.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Maintainer: Matthew Poremba <matthew.poremba@amd.com>
For non-KVM CPUs the VBIOS memory falls into an I/O hole and therefore
gets routed to the PIO bus in gem5. This gets routed to the GPU in the
case of a ROM write. We write to the ROM as a way to "load" the VBIOS
without creating holes in the KVM VM.
This write method allows the same scripts as KVM to be used by writing
to the ROM area and overwriting what might already be there from the
--gpu-rom option.
Change-Id: I8c2d2aa05a823569a774dfdd3bf2d2e773f38683
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/70037
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
The GPUFS checkpoint restoration mechanism expects to find a
PM4MapQueues packet in the checkpoint. Since this was not being
checkpointed, the restore phase retrieved a null packet which led to a
segmentation fault. This commit adds PM4MapQueues to the checkpoint and
restores it when deserializing the checkpoint
Change-Id: Ib74a9f36fe89d740a74f94314ada41ecc363abe9
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/69298
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matthew Poremba <matthew.poremba@amd.com>
