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
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>
Currently when RLC queues (user mode queues) are mapped, the read/write
pointers of the ring buffer are set to zero. However, these queues could
be unmapped and then remapped later. In that situation the read/write
pointers should be the previous value before unmapping occurred. Since
the read pointer gets reset to zero, the queue begins reading from the
start of the ring, which usually contains older packets. There is a 99%
chance those packets contain addresses which are no longer in the page
tables which will cause a page fault.
To fix this we update the MQD with the current read/write pointer values
and then writeback the MQD to memory when the queue is unmapped. This
requires adding a pointer to the MQD and the host address of the MQD
where it should be written back to. The interface for registering RLC
queue is also simplified. Since we need to pass the MQD anyway, we can
get values from it as well.
Fixes b+tree and streamcluster from rodinia (when using RLC queues).
Change-Id: Ie5dad4d7d90ea240c3e9f0cddf3e844a3cd34c4f
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/65791
Tested-by: kokoro <noreply+kokoro@google.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Currently the SDMA queue type is guessed in the trap method by looking
at which queue in the engine is processing packets. It is possible for
both queues to be processing (e.g., one queue sent a DMA and is waiting
then switch to another queue), triggering an assert.
Instead store the queue type in the queue itself and use that type in
trap to determine which ring ID to use for the interrupt packet.
Change-Id: If91c458e60a03f2013c0dc42bab0b1673e3dbd84
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/65691
Maintainer: Jason Lowe-Power <power.jg@gmail.com>
Reviewed-by: Jason Lowe-Power <power.jg@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
The current SDMA wrap around handling only considers the ring buffer
location as seen by the GPU. Eventually when the end of the SDMA ring
buffer is reached, the driver waits until the rptr written back to the
host catches up to what the driver sees before wrapping around back to
the beginning of the buffer. This writeback currently does not happen at
all, causing hangs for applications with a lot of SDMA commands.
This changeset first fixes the sizes of the queues, especially RLC
queues, so that the wrap around occurs in the correct place. Second, we
now store the rptr writeback address and the absoluate (unwrapped) rptr
value in each SDMA queue. The absolulte rptr is what the driver sends to
the device and what it expects to be written back.
This was tested with an application which basically does a few hundred
thousand hipMemcpy() calls in a loop. It should also fix the issue with
pannotia BC in fullsystem mode.
Change-Id: I53ebdcc6b02fb4eb4da435c9a509544066a97069
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/65351
Maintainer: Jason Lowe-Power <power.jg@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Jason Lowe-Power <power.jg@gmail.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
SDMA traps are used in the driver as a DMA fence. To pass a fence, the
SDMA sends the driver the interrupt context from a trap packet and the
ring ID which specifies which queue in the SDMA engine is passing a
fence. Currently the interrupt context is using the wrong value in the
packet and the ring ID is hard-coded to always be the gfx queue.
This changeset uses the correct interrupt context from the SDMA packet
and sets the ring ID to either 0 if the gfx queue is currently being
processed or 3 if the page queue is being processed.
The relevant interrupt service routine in the driver can be found at:
https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver/blob/roc-4.3.x/
drivers/gpu/drm/amd/amdgpu/sdma_v4_0.c#L2129
Change-Id: Ie4a4a9d6ab1d3bf83bf76bb57a02a91100217b51
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/65093
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
The current implementation of SDMA copy calls the GPU memory manager's
read/write method one time passing a physical address as the
source/destination. This implicitly assumes the physical addresses are
contiguous which is generally not true for large allocations. This
results in reading from/writing to the wrong address.
This changeset fixes the problem by copying large copies in chunks of
the minimum possible page size on the GPU (4kB). Each page is translated
seperately to ensure the correct physical address. The final copy "done"
callback is only used for the last transfer. The transfers should
complete in order so the copy command will not complete until all chunks
have been copied. Tested and verified on an application with a large
allocation (~5GB).
Change-Id: I27018a963da7133f5e49dec13b0475c3637c8765
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/64752
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
This debug flag is used to print spammy SDMA DPRINTFs, such as an SDMA
copy printing the data of large transfers 8 bytes per line at a time. For
those prints, the SDMAEngine flag will now only print the first and last
qword of the transfer and the new SDMAData flag is needed for verbose
data printing. This makes the SDMAEngine flag still useful for verifying
copies in applications with predictable data such as square.
Additionally, the memory allocation/deallocation done solely for a print
statement is removed in favor of casting the data to the printed type.
Change-Id: I18c1918ef9085cca4570f79881ee63d510ccc32f
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/64452
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
SDMA atomic packets are used in conjunction with RLC queues in SDMA for
synchronization similar to how HSA signals are used with BLIT kernels
when SDMA is disabled. Implement a skeleton of the SDMA atomic packet
methods as well as the atomic add64 operation.
The atomic add operation appears to be the only operation used in ROCm,
so this implementation is fairly complete. See:
https://github.com/RadeonOpenCompute/ROCR-Runtime/blob/
rocm-4.2.x/src/core/runtime/amd_blit_sdma.cpp#L880
Change-Id: I62cc337f2ffe590bdb947b48053760ee8b3a6f32
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/63174
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
There can only ever be two RLC queues maximum. Use this information for
a simpler data structure to store doorbell information. The patch
changes the std::unordered_map previously used to std::array. This will
also be useful in avoiding erase-while-iterating issues needed to
unregister all queues at once.
Change-Id: I95600e40de51cb1a992a20bcebaf7580ea4d0be8
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/63172
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
SDMA write, copy, and ptePde use GPUMemMgr to write to device memory and
were dangerously not waiting for write completion which could result in
data not being completely written to memory, the data buffer being freed
and potentially reused in the simulator, or advancing to the next SDMA
packet before the previous one is complete.
This changeset adds callbacks for the corresponding "done" methods
similar to what the dmaVirt methods call when reading or writing to host
memory to fix this issue.
Change-Id: I44ce14c13f812ea2a7a76438e12a6ed7c6e0bff0
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/62715
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Jason Lowe-Power <power.jg@gmail.com>
Reviewed-by: Jason Lowe-Power <power.jg@gmail.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Now that the memory manager can DMA read from device memory, allow the
linear copy SDMA packet to use device memory as a source. This is used
when copying memory from device to host when SDMA engines are enabled.
This improves simulation performance over using (simulated) BLIT kernels
with SDMA engines disabled.
Change-Id: I1f41b294022f0049d154a401c1dc885abb4f223b
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/62713
Maintainer: Jason Lowe-Power <power.jg@gmail.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Reviewed-by: Jason Lowe-Power <power.jg@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
The PM4 packet processor is handling all non-HSA GPU packets such
as packets for (un)mapping HSA queues. This commit pulls many
Linux structs and defines out into their own files for clarity.
Finally, it implements the VMID related functions in AMDGPU device.
Change-Id: I5f0057209305404df58aff2c4cd07762d1a31690
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/53068
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
SDMAEngine handles copies to device memory. This commit
updates sdma_packets.hh style as well. Added several methods needed by
SDMAEngine to GPU device including GART table, various getters, and
aperture range checkers. Move the MMIO interface from GPUController to
SDMAEngine. Create an SDMA MMIO and commands header with only the macros
we use so that we don't need to check in multi-thousand line header
files from the linux kernel. Keep SOC15 IH client ID macros as that file
is small.
Change-Id: I986fede90cc1bc16ee56d4e8598cf9283bde034e
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/53064
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
