This was found while comparing a diverging execution against QEMU traces
and checking for the first mismatched program counter. Fortunately this
was
caused by a branch shortly after this incorrect computation but still
took
a long time to track down.
There are two issues here: the decoder had inverted the cases for *S and
*A,
and the sign bit was wrong for VFN*.
This PR has 3 commits:
- Update scaling methods to scale by multiplication or division when
upcasting or downcasting respectively.
- Preserve the sign when a microscaling conversion results in NaN or
infinity to match hardware.
- Rework rounding to handle cases where conversion results in a denormal
number in the output type so that the value is correct.
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
The current implementation does not correctly convert subnormal numbers
(number that fill the underflow gap around zero in floating-point
arithmetic). This commit reworks the rounding code to get correct
results.
First, the min_exp is set to 0 which allows for numbers to become
subnormal when rounding. Second, the rounding code now uses something
closer to "GRS" rounding (guard, round, sticky) which represent the
first bit removed when rounding to a smaller type, the next second bit
removed, and whether any of the other bits removed are one. More details
can be found in the code comments.
Change-Id: Idcd2f1e4383e4012fc3abf73b1f73c847d44f67b
The implementation of microscaling formats uses the Open Compute Project
specification which includes a sign bit for NaN and infinity. This
should be preserved when a conversion results in NaN or infinity.
Change-Id: Id9e99324c6486e256c699016aff301d5f06814d5
Currently there is only a scale() method which multiplies a microscaling
type by an int8 value. This should only be applied when upcasting to
a larger type after conversion to match hardware. When downcasting to a
smaller type, the scaling method should divide by the int8 value before
conversion.
This commit adds both scaling methods.
Change-Id: Ibafa8caa389cde4df609e536cd53bd2289959420
At the moment, a hart does not halt if there are pending interrupts.
However, an implementation can also consider the enable status of the
individual interrupts, i.e., a halted hart would only resume if there
are locally enabled pending interrupts. This commit introduces this
behavior. The wfi behavior is controlled by the new configuration
variable wfi_pending_resume of RiscvISA.
Change-Id: I316239f9732c6e73e6ad692491bca08d773dd995
---------
Signed-off-by: Robert Hauser <robert.hauser@uni-rostock.de>
In gem5, we use the same code base for RISC-V 32 and 64.
However, if we need to allow modifiable XLEN control on CSR.mstatus in
the future, we should follow the RISC-V ISA manual to sign-extend all
the register results, including PC and GPR. If this feature implemented,
the simulator needs to handle user-mode in RV32 but CSR.SATP sets to
Sv39. In this case, 0x80000000 and 0xffffffff80000000 are different
addresses in the 64-bit S-Mode perspective, but they are the same in the
32-bit U-Mode perspective. We should avoid this wrong behavior happening
before we implement this feature.
Thus, we need to sign-extend the results of all the addresses, including
the PC and memory addresses, which currently use zero-extend. As
specified in the RISC-V ISA manual, we use zero-extend in narrow XLEN
mode for the physical address implemented in TLB.
Changes based on spec:
1. Sign-extend narrow XLEN:
https://github.com/riscv/riscv-isa-manual/blob/riscv-isa-release-b7a445a-2024-07-02/src/machine.adoc?plain=1#L567
2. Zero-extend physical address:
https://github.com/riscv/riscv-isa-manual/blob/riscv-isa-release-b7a445a-2024-07-02/src/supervisor.adoc?plain=1#L1670
Signed-off-by: Yangyu Chen <cyy@cyyself.name>
Implementing generic reset method for MMU allows each ISA implementing
their own reset methods. The default reset MMU method is flush all TLB
entries. For example, The RISC-V needs to do PMP reset when received the
reset signal, but the TLBs don't require to be flushed.
Change-Id: I158261570fb6e5216ec105fbdc53460f83f88d15
This makes it easier to debug unexpected semihosting outputs (in my case
a wrong buffer argument was being passed).
Change-Id: I342610a92fb8efe121d030f7b9ea3307efc4fec3
These registers were only handled in AArch64 mode but are also
accessible as a c14 registers for AArch32.
Change-Id: I62fe54427e96265df0589308afa1b5d665dbf210
In AArch32 mode it is possible to read a 64-bit counter using mrrc.
Instead of truncating in the PMU code, just allow the instruction
implementation to truncate to 32 bits if accessed using mrc.
Change-Id: I77620f6d1852a7d9e79c1ecee50f4297b4103b1c
If SDWAB or DPP are used on a VOPC instruction and those are not
implemented, it is highly likely to be a problem for the application.
Rather than continue to execute and cause undefined behavior, exit the
simulation with a panic showing the line of the instruction causing the
issue.
Change-Id: Ib3f94df7445d068b26907470c1f733be16cd2fc2
Add a panic if SDWA or DPP is used for an instruction which does not
implement support for it. If an application uses SDWA or DPP it likely
does not operate in the same way as the base instruction and therefore
gem5 should panic rather than continue. It is likely data is incorrect
which will make it more difficult to debug an application.
Change-Id: I68ac448b0d62941761ef4efa0169f95796270f48
This shows an example of how to use the previous commit which adds an
SDWAB helper. The execute() method of both are the same with the
exception of the lambda function passed to the helper method.
Change-Id: I5ffe361440b4020b9f7669c0ed946aa6b3bbec25
Implement a SDWAB helper which accepts a dynamic instruction and a
lambda function defining a comparison function taking two values and
returning a comparison result of 0 or 1 for false or true.
Current instructions which implement SDWA do so on a per-instruction
basis which adds a lot of redundant code. This allows for generic SDWAB
implementations for VOPC instructions.
All modifiers are implemented assuming that SDWBA VOPC instruction
comparison types may be U32, I32, F32, U16, I16, F16 (which exist) but
is extendible to I8, U8, or F8.
Change-Id: Idab58a327c29dd19a1a5457237f3799a04f2031b
Some instructions are clamping floating point outputs unconditionally,
leading to incorrect results. This commit finds instructions with this
issue and checks the clamp bit before applying clamp.
Change-Id: Ibc6de3813d81fd4f9d2c98dd497d19dd34cf6bde
This PR improves the vector register groups overlap check in
widening/narrowing
instructions.
- Fix wrong illegal overlap condition between VS2 and VD vector register
groups.
- Also check VS1 vector register group for overlap with VD in
vector-vector
instructions.
- Parametrize widening/narrowing factors in overlap check function to
potentially
handle more cases.
Fixes issue #442.
This fixes the remaining implicit int/float conversions and enables the
float conversion warnings for clang when building the Arm instruction
execution logic. This depends on the previous fixes.
Change-Id: I51aac94644a483175842c36da2d49d308aaceb49
TimingExpression enables runtime calculation of the commit latency in
MinorCPU. For this, machInst is obtained by getEMI() to match it with a
given instruction. At default, getEMI() always returns 0 and is
therefore overwritten to enable timing expressions for RISC-V. This was
already done for ARM (see src/arch/arm/insts/static_inst.hh).
Change-Id: I03d669b3439fd24e00cbf893f5db9951dfe56b1f
Signed-off-by: Robert Hauser <robert.hauser@uni-rostock.de>
Updated the bib information of the local RISC-V interrupts.
Change-Id: I666c3df4529e159bd1946ca1a9623e47f84d5d9e
Signed-off-by: Robert Hauser <robert.hauser@uni-rostock.de>
Currently, only the VS2 register group is checked for overlap with VD
when executing a widening/narrowing instruction. This commits extends
the check to VS1, when applicable (i.e. vector-vector operations).
Change-Id: I892b7717c01e25546fb41e05afbd08fc40c60c59
As of now, the widening/narrowing vector register groups overlap check
always assumes a SEW multiplication factor equal to 2 (for either VD or
VS2). This commits aims at making this check more generic.
Change-Id: I4311fc3624cd324ccfdf2a1920a19efc85357120
This commit fixes the overlap check between VS2 and VD register groups
in vector widening instructions. While the narrowing instructions check
is correct, the widening one has to differentiate between two cases
(Vs2 EEW = 2*SEW and Vs2 EEW = SEW). In the first case, overlap is
allowed, as the EEW is the same as Vd. In the second case, the overlap
legality check has to be adapted to use the Vs2 EMUL to calculate the
boundaries. The rule has been derived again from Section 5.2 of RISC-V
"V" Vector Extension specifications, version 1.0.
The patch also includes some small code refactoring, e.g. using
already defined vlmul and constants for vector operands.
Fixes issue #442.
Change-Id: Ic87095fb9079e6c8f53b9a0d79fbf531a85dc71d
Vector reduce float (widening and non-widening) and integer (widening)
instructions initialize the reduce loop operation with the first element
of the destination register (i.e. `Vd[0]`).
Since all reductions per spec seem to be `Vd[0] = Vs1[0] + Vs2[*]`
(where `+` is an arbitrary binary op and `*` indicates all active
elements) gem5 will calculate this incorrectly if `Vd[0]` and/or
`Vs1[0]` are non-neutral for the operation (the later case being because
it's not taken into account at all).
To solve this we just have to initialize the reduction loop to `Vs1[0]`
(the non-widening integer reduction already does this).
The gem5 crashed when user try to update register value from GDB because
PR[1] changes the index of CSR_XSTATUS to MISCREG_XSTATUS, which is out
of NUM_PHYS_MISCREGS.
The CSR_XSTATUS should use setRegWithMask to update it.
[1] : https://github.com/gem5/gem5/pull/1099
gem5 issue: https://github.com/gem5/gem5/issues/1299
Change-Id: Iefc0d1f5adfb98ecfda0e74907964b47d1864b6d
These two commits add agnostic capability for both tail/mask policies,
for vector memory and arithmetic instructions respectively. The common
policy for instructions is to act as undisturbed if one is (i.e. tail or
mask), or write all 1s if none.
For those instructions in which multiple micro instructions are
instantiated to write to the same register (`VlStride` and `VlIndex` for
memory, and `VectorGather`, `VectorSlideUp` and `VectorSlideDown` for
arithmetic), a (new) micro instruction named `VPinVdCpyVsMicroInst` has
been used to pin the destination register so that there's no need to
copy the partial results between them. This idea is similar to what's on
ARM's SVE code. This micro also implements the tail/mask policy for this
cases.
Finally, it's worth noting that while now using an agnostic policy for
both tail/mask should remove all dependencies with old destination
registers, there's an exception with `VectorSlideUp`. The
`vslideup_{vx,vi}` instructions need the elements in the offset to be
unchanged. The current implementation overrides the current vta/vma and
makes them act as undisturbed, since they require the old destination
register anyways. There's a minor issue with this though, as
`v{,f}slide1up` variants do not need this, but since they share the same
constructor, will act all the same.
Related issue #997.
Implement small translation table extension.
This feature relaxes the lower limit on the size of the translation
tables, by increasing the maximum permitted values of the T1SZ and T0SZ
field in: TCR_EL1, TCR_EL2, TCR_EL3,VTCR_EL2 and VSTCR_EL2
Change-Id: I4c2187815b2d7f14407edb38095c6bcc2004b62a
Signed-off-by: Giacomo Travaglini <giacomo.travaglini@arm.com>
We are currently using the LongDecriptor for both stage1
and stage2 translations. There are several cases where
the bitfield meaning changes depending on the translation
stage.
Change-Id: Ic33d9ef225a57fd79ce2b4bf47896aeb6bdd8d9c
Signed-off-by: Giacomo Travaglini <giacomo.travaglini@arm.com>
For ARMv8 CPUs this register allows reading a 64-bit cycle counter in
from 32-bit execution state.
Change-Id: I7cd9e2711ada5156920440cc3c89e7a74ca54a49
This patch is adding a functional implementation of FEAT_XS. Unless we
operate with DVM enabled, TLBIs broadcasting is accomplished in 0 time;
so there is no timing benefit introduced by enabling FEAT_XS other than
the way it affects TLB management (invalidation)
Change-Id: I067cb8b7702c59c40c9bbb8da536a0b7f3337b5d
Signed-off-by: Giacomo Travaglini <giacomo.travaglini@arm.com>
The acquire-release flavor of the ldadd instruction should read ldaddalx
(eg. ldaddalb/ldaddalh) according to specification. However, this is
currently noted as ldadd"la"x (eg. ldaddlab/ldaddlah).
Issue: https://github.com/gem5/gem5/issues/1224
Change-Id: Ib932fa0e572207729c923c27f24c34cc21dff0e5
Co-authored-by: Bobby R. Bruce <bbruce@ucdavis.edu>
Previously, all of the TLB lookup/insert functions were using the full
virtual addresses even though the variables in the functions said "vpn."
This change explicitly converts the virtual address to the VPN without
any least significant zeros for the offset. I.e., vpn >> page_size.
The main bug solved in this changeset is the asid was |'d with the upper
bits of the virtual address, but sometimes there were all 1's.
Therefore, you could get a TLB hit even if the ASID was different.
Interestingly, the page that seemed to cause these issues was a 1 GiB
page.
This change also starts refactoring some of the page table details to
support sv46 and sv57 page table formats.
In my testing, the Linux kernel boot uses large pages (even OpenSBI uses
large pages), so it seems that large pages also work. However, this
seems like magic to me, so I'm not sure if it's correct.
This change also updates some asserts, and debug statements with more
useful debugging information.
Partially fixes#1235. More testing needs to be done to be confident.
Currently, gem5's inst tracer prints the whole vector register container
by default. The size of vector register containers in gem5 is the
maximum size allowed by the ISA. For vector-length agnostic (VLA) vector
registers, this means ARM SVE vector container is 2048 bits long, and
RISC-V vector container is 65535 bits long. Note that VLA implementation
in gem5 allows the vector length to be varied within the limit specified
by the ISAs.
However, in most use cases of gem5, the vector length is much less than
65535 bits. This causes two issues: (1) the vector container requires
allocating and moving around a large amount of unused data while only a
fraction of it is used, and (2) printing the execution trace of a vector
register results in a wall of text with a small amount of useful data.
This change addresses the problem (2) by providing a mechanism to limit
the amount data printed by the instruction tracer. This is done by
adding a function printing the first X bits of a vector register
container, where X is the vector length determined at runtime, as
opposed to the vector container size, which is determined at compilation
time.
Change-Id: I815fa5aa738373510afcfb0d544a5b19c40dc0c7
---------
Signed-off-by: Hoa Nguyen <hn@hnpl.org>
This is a follow-up on the discussion here [1].
The IsInvalid flag was previously defined as an instruction that does
not appear in the ISA. However, a micro-architecture can choose to not
recognize an instruction in and raise illegal instruction fault even if
the instruction is in the ISA.
This change modifies the definition of a Invalid instruction such that,
if a StaticInst instruction is marked as IsInvalid, it means the
instruction is not recognized by the decoder. This means that any
instruction recognized by the decoder are not invalid, even if the
instruction is not in the official ISA spec; e.g., m5
pseudo-instructions.
Note that instructions that are recognized by the decoder but are chosen
to act as a nop are not invalid. This applies to WarnUnimplemented
instructions, e.g. hint instructions.
[1] https://github.com/gem5/gem5/pull/1071
Change-Id: I1371b222d8b06793d47f434d0f148c5571672068
Signed-off-by: Hoa Nguyen <hn@hnpl.org>
For scratch instructions only, this bit specifies if an offset in a VGPR
should be used for address calculation. This is new in MI300 and was
previously the LDS bit. The LDS bit is rarely used and in fact gem5 does
not even check this bit.
This fixes a bug when SADDR == 0x7f (i.e., no SGPR should be used) where
a VGPR was being added to the address when it should have been ignored.
Change-Id: I9864379692df6795b25b58b98825da05d18fc5db
