Scratch memory requests that are larger than one dword are using a
different memory layout than global instructions. Rather than being
placed contiguously, each dword is interleaved 64 lanes * 4 bytes away
as described in Section 9.1.5.2. "Swizzled Buffer Addressing" in the
MI300 specification. This was verified by comparing MI300 output (which
uses scratch_ instructions) with MI200 (which uses buffer instructions).
MI300 FashionMNIST bs=1 now matches CPU reference.
This requires several changes to the instruction implementations:
- For stores, data in the GPUDynInst can be swizzled before the data is
written to memory. This is easy to do using a helper method. This is
done in the template<int N> variant of initMemWrite. To use this x2
stores are changed to use template<int N> rather than loading a U64. The
swizzle function is renamed to swizzleAddr to avoid confusion with
swizzleData.
- For loads, data is unswizzled in completeAcc when writing register
values. This is not as easy to implement as a helper and is thus
implemented for the three load instructions that load more than one
dword.
- Accessing swizzled data requires at least one packet per dword. A new
GPU memory helper is added to create these packets for scratch requests
specifically. This is called in the template<int N> variant of
initMemRead / initMemWrite. Loads and stores of x2 are changed to use
this variant instead of accessing a U64.
The GPUDynInst status vector restrictions are increased to allow for
swizzled x4 accesses. For simplicity this does not currently support
misaligned swizzled accesses and will panic upon seeing such a case.
Change-Id: Ic686c51e28e0af029a043d5a5b3d4069f2cb94f9
Memory instructions acquire coalescer tokens in the schedule stage.
Currently this is only done for buffer and flat instructions, but not
flat global or flat scratch. This change now acquires tokens for flat
global and flat scratch instructions. This provides back-pressure to the
CUs and helps to avoid deadlocks in Ruby.
The change also handles returning tokens for buffer, flat global, and
flat scratch instructions. This was previously only being done for
normal flat instructions leading to deadlocks in some applications when
the tokens were exhausted.
To simplify the logic, added a needsToken() method to GPUDynInst which
return if the instruction is buffer or any flat segment.
The waitcnts were also incorrect for flat global and flat scratch. We
should always decrement vmem and exp count for stores and only normal
flat instructions should decrement lgkm. Currently vmem/exp are not
decremented for flat global and flat scratch which can lead to deadlock.
This change set fixes this by always decrementing vmem/exp and lgkm only
for normal flat instructions.
Change-Id: I673f4ac6121e4b5a5e8491bc9130c6d825d95fc5
Flat scratch instructions (aka private) are the 3rd and final segment of
flat instructions in gfx9 (Vega) and beyond. These are used for things
like spills/fills and thread local storage. This commit enables two
forms of flat scratch instructions: (1) flat_load/flat_store
instructions where the memory address resolves to private memory and (2)
the new scratch_load/scratch_store instructions in Vega. The first are
similar to older generation ISAs where the aperture is unknown until
address translation. The second are instructions guaranteed to go to
private memory.
Since these are very similar to flat global instructions there are
minimal changes needed:
- Ensure a flat instruction is either regular flat, global, XOR scratch
- Rename the global op_encoding methods to GlobalScratch to indicate
they are for both and are intentionally used.
- Flat instructions in segment 1 output scratch_ in the disassembly
- Flat instruction executed as private use similar mem helpers as global
- Flat scratch cannot be an atomic
This was tested using a modified version of the 'square' application:
template <typename T>
__global__ void
scratch_square(T *C_d, T *A_d, size_t N)
{
size_t offset = (blockIdx.x * blockDim.x + threadIdx.x);
size_t stride = blockDim.x * gridDim.x ;
volatile int foo; // Volatile ensures scratch / unoptimized code
for (size_t i=offset; i<N; i+=stride) {
foo = A_d[i];
C_d[i] = foo * foo;
}
}
Change-Id: Icc91a7f67836fa3e759fefe7c1c3f6851528ae7d
The amdgpu driver supports reading and writing scalar and vector memory
addresses that reside in system memory. This is commonly used for things
like blit kernels that perform host-to-device or device-to-host copies
using GPU load/store instructions.
This is done by utilizing the system hub device added in a prior
changeset. Memory packets translated by the Scalar or VMEM TLBs will
have the correspoding system request field set from the PTE in the TLB
which can be used in the compute unit to determine if a request is for
system memory or not.
Another important change is to return global memory tokens for system
requests. Since these do not flow through the GPU coalescer where the
token is returned, the token can be returned once the request is known
to be a system request.
Change-Id: I35030e0b3698f10c63a397f96b81267271e3130e
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/57711
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Remove the line "For use for simulation and test purposes only" in files
were AMD is the only copyright holder listed in the header. This happens
to be the case for all files where this line exists, removing it
completely from gem5.
Change-Id: I623f266b002f564301b28774f49081099cfc60fd
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/53943
Reviewed-by: Jason Lowe-Power <power.jg@gmail.com>
Maintainer: Jason Lowe-Power <power.jg@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Global instructions are new in Vega and are essentially FLAT
instructions from GCN3 but guaranteed to go to global memory where as
flat can go to global or local memory.
This reworks the flat instruction classes so that the initiateAcc /
execute / completeAcc logic can be reused for flat, global, and later
scratch subtypes of flat instructions. The decoder creates a flat
instruction class which sets instruction flags based on the flat
instruction's SEG field. There are new initOperandInfo and
generateDissasmbly methods for flat and global. The number of operands
and operand index getters are modified to check the flags and return the
correct value for the subtype.
Change-Id: I1db4a3742aeec62424189e54c38c59d6b1a8d3c1
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/47106
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Reviewed-by: Kyle Roarty <kyleroarty1716@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Apply the gem5 namespace to the codebase.
Some anonymous namespaces could theoretically be removed,
but since this change's main goal was to keep conflicts
at a minimum, it was decided not to modify much the
general shape of the files.
A few missing comments of the form "// namespace X" that
occurred before the newly added "} // namespace gem5"
have been added for consistency.
std out should not be included in the gem5 namespace, so
they weren't.
ProtoMessage has not been included in the gem5 namespace,
since I'm not familiar with how proto works.
Regarding the SystemC files, although they belong to gem5,
they actually perform integration between gem5 and SystemC;
therefore, it deserved its own separate namespace.
Files that are automatically generated have been included
in the gem5 namespace.
The .isa files currently are limited to a single namespace.
This limitation should be later removed to make it easier
to accomodate a better API.
Regarding the files in util, gem5:: was prepended where
suitable. Notice that this patch was tested as much as
possible given that most of these were already not
previously compiling.
Change-Id: Ia53d404ec79c46edaa98f654e23bc3b0e179fe2d
Signed-off-by: Daniel R. Carvalho <odanrc@yahoo.com.br>
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/46323
Maintainer: Bobby R. Bruce <bbruce@ucdavis.edu>
Reviewed-by: Bobby R. Bruce <bbruce@ucdavis.edu>
Reviewed-by: Matthew Poremba <matthew.poremba@amd.com>
Tested-by: kokoro <noreply+kokoro@google.com>
This commit removes functions that indexed into the
vectors that held the operands. Instead, for-each loops
are used, iterating through one of 6 vectors
(src, dst, srcScalar, srcVec, dstScalar, dstVec)
that all hold various (potentially overlapping)
combinations of the operands.
Change-Id: Ia3a857c8f6675be86c51ba2f77e3d85bfea9ffdb
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/42212
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
This change removes the GPUDynInstPtr argument from
getRegisterIndex(). The dynamic inst was only needed
to get access to its parent WF's state so it could
determine the number of scalar registers the wave was
allocated. However, we can simply pass the number of
scalar registers directly. This cuts down on shared
pointer usage.
Change-Id: I29ab8d9a3de1f8b82b820ef421fc653284567c65
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/42210
Tested-by: kokoro <noreply+kokoro@google.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Matt Sinclair <mattdsinclair@gmail.com>
Previously, with HSAIL, we were guaranteed by the HSA specification
that the GPU will never issue unaligned accesses. However, now
that we are directly running GCN this is no longer true.
Accordingly, this commit adds support for unaligned accesses.
Moreover, to reduce the replication of nearly identical
code for the different request types, I also added new helper
functions that are called by all the different memory request
producing instruction types in op_encodings.hh.
Adding support for unaligned instructions requires changing
the statusBitVector used to track the status of the memory
requests for each lane from a bit per lane to an int per lane.
This is necessary because an unaligned access may span multiple
cache lines. In the worst case, each lane may span multiple
cache lines. There are corresponding changes in the files that
use the statusBitVector.
Change-Id: I319bf2f0f644083e98ca546d2bfe68cf87a5f967
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/29920
Reviewed-by: Anthony Gutierrez <anthony.gutierrez@amd.com>
Reviewed-by: Matt Sinclair <mattdsinclair@gmail.com>
Maintainer: Anthony Gutierrez <anthony.gutierrez@amd.com>
Tested-by: kokoro <noreply+kokoro@google.com>
Neither assert(0) nor assert(false) give any hint as to why control
getting to them is bad, and their more descriptive versions,
assert(0 && "description") and assert(false && "description"), jury
rig assert to add an error message when the utility function panic()
already does that directly with better formatting options.
This change replaces that flavor of call to assert with panic, except
in the actual code which processes the formatting that panic uses (to
avoid infinitely recurring error handling), and in some *.sm files
since I don't know what rules those have to follow and don't want to
accidentaly break them.
Change-Id: I8addfbfaf77eaed94ec8191f2ae4efb477cefdd0
Reviewed-on: https://gem5-review.googlesource.com/c/14636
Reviewed-by: Brandon Potter <Brandon.Potter@amd.com>
Reviewed-by: Jason Lowe-Power <jason@lowepower.com>
Maintainer: Jason Lowe-Power <jason@lowepower.com>
The gpu_dyn_inst.hh file was missing a clone method from
inherited classes. (The clone method is the way to implement
the prototype design pattern.) Because the inherited clone
method was declare as pure virtual, the method needed to
be implemented. Otherwise, the compiler complains that the
class is abstract.
Change-Id: I38782d5f7379f32be886401f7c127fe60d2f8811
Reviewed-on: https://gem5-review.googlesource.com/12108
Reviewed-by: Jason Lowe-Power <jason@lowepower.com>
Reviewed-by: Anthony Gutierrez <anthony.gutierrez@amd.com>
Maintainer: Anthony Gutierrez <anthony.gutierrez@amd.com>
for HSAIL an operand's indices into the register files may be calculated
trivially, because the operands are always read from a register file, or are
an immediate.
for machine ISA, however, an op selector may specify special registers, or
may specify special SGPRs with an alias op selector value. the location of
some of the special registers values are dependent on the size of the RF
in some cases. here we add a way for the underlying getRegisterIndex()
method to know about the size of the RFs, so that it may find the relative
positions of the special register values.
we are removing doGmReturn from the GM pipe, and adding completeAcc()
implementations for the HSAIL mem ops. the behavior in doGmReturn is
dependent on HSAIL and HSAIL mem ops, however the completion phase
of memory ops in machine ISA can be very different, even amongst individual
machine ISA mem ops. so we remove this functionality from the pipeline and
allow it to be implemented by the individual instructions.
this patch removes the GPUStaticInst enums that were defined in GPU.py.
instead, a simple set of attribute flags that can be set in the base
instruction class are used. this will help unify the attributes of HSAIL
and machine ISA instructions within the model itself.
because the static instrution now carries the attributes, a GPUDynInst
must carry a pointer to a valid GPUStaticInst so a new static kernel launch
instruction is added, which carries the attributes needed to perform a
the kernel launch.
Eliminate the VSZ constant that defined the Wavefront size (in numbers of work
items); replaced it with a parameter in the GPU.py configuration script.
Changed all data structures dependent on the Wavefront size to be dynamically
sized. Legal values of Wavefront size are 16, 32, 64 for now and checked at
initialization time.
