Currently, we use `AAAMDWavesPerEU` to iteratively update values based
on attributes from the associated function, potentially propagating
user-annotated values, along with `AAAMDFlatWorkGroupSize`. Similarly,
we have `AAAMDFlatWorkGroupSize`. However, since the value calculated
through the flat workgroup size always dominates the user annotation
(i.e., the attribute), running `AAAMDWavesPerEU` iteratively is
unnecessary if no user-annotated value exists.
This PR completely rewrites how the `amdgpu-waves-per-eu` attribute is
handled in `AMDGPUAttributor`. The key changes are as follows:
- `AAAMDFlatWorkGroupSize` remains unchanged.
- `AAAMDWavesPerEU` now only propagates user-annotated values.
- A new function is added to check and update `amdgpu-waves-per-eu`
based on the following rules:
- No waves per eu, no flat workgroup size: Assume a flat workgroup size
of `1,1024` and compute waves per eu based on this.
- No waves per eu, flat workgroup size exists: Use the provided flat
workgroup size to compute waves-per-eu.
- Waves per eu exists, no flat workgroup size: This is a tricky case. In
this PR, we assume a flat workgroup size of `1,1024`, but this can be
adjusted if a different approach is preferred. Alternatively, we could
directly use the user-annotated value.
- Both waves per eu and flat workgroup size exist: If there’s a
conflict, the value derived from the flat workgroup size takes
precedence over waves per eu.
This PR also updates the logic for merging two waves per eu pairs. The
current implementation, which uses `clampStateAndIndicateChange` to
compute a union, might not be ideal. If we think from ensure proper
resource allocation perspective, for instance, if one pair specifies a
minimum of 2 waves per eu, and another specifies a minimum of 4, we
should guarantee that 4 waves per eu can be supported, as failing to do
so could result in excessive resource allocation per wave. A similar
principle applies to the upper bound. Thus, the PR uses the following
approach for merging two pairs, `lo_a,up_a` and `lo_b,up_b`: `max(lo_a,
lo_b), max(up_a, up_b)`. This ensures that resource allocation adheres
to the stricter constraints from both inputs.
Fix#123092.
The default maximum waves/EU returned by the family of
`AMDGPUSubtarget::getWavesPerEU` is currently the maximum number of
waves/EU supported by the subtarget (only a valid occupancy range in
"amdgpu-waves-per-eu" may lower that maximum). This ignores maximum
achievable occupancy imposed by flat workgroup size and LDS usage,
resulting in situations where `AMDGPUSubtarget::getWavesPerEU` produces
a maximum higher than the one from
`AMDGPUSubtarget::getOccupancyWithWorkGroupSizes`.
This limits the waves/EU range's maximum to the maximum achievable
occupancy derived from flat workgroup sizes and LDS usage. This only has
an impact on functions which restrict flat workgroup size with
"amdgpu-flat-work-group-size", since the default range of flat workgroup
sizes achieves the maximum number of waves/EU supported by the
subtarget.
Improvements to the handling of "amdgpu-waves-per-eu" are left for a
follow up PR (e.g., I think the attribute should be able to lower the
full range of waves/EU produced by these methods).
Occupancy (i.e., the number of waves per EU) depends, in addition to
register usage, on per-workgroup LDS usage as well as on the range of
possible workgroup sizes. Mirroring the latter, occupancy should
therefore be expressed as a range since different group sizes generally
yield different achievable occupancies.
`getOccupancyWithLocalMemSize` currently returns a scalar occupancy
based on the maximum workgroup size and LDS usage. With respect to the
workgroup size range, this scalar can be the minimum, the maximum, or
neither of the two of the range of achievable occupancies. This commit
fixes the function by making it compute and return the range of
achievable occupancies w.r.t. workgroup size and LDS usage; it also
renames it to `getOccupancyWithWorkGroupSizes` since it is the range of
workgroup sizes that produces the range of achievable occupancies.
Computing the achievable occupancy range is surprisingly involved.
Minimum/maximum workgroup sizes do not necessarily yield maximum/minimum
occupancies i.e., sometimes workgroup sizes inside the range yield the
occupancy bounds. The implementation finds these sizes in constant time;
heavy documentation explains the rationale behind the sometimes
relatively obscure calculations.
As a justifying example, consider a target with 10 waves / EU, 4 EUs/CU,
64-wide waves. Also consider a function with no LDS usage and a flat
workgroup size range of [513,1024].
- A group of 513 items requires 9 waves per group. Only 4 groups made up
of 9 waves each can fit fully on a CU at any given time, for a total of
36 waves on the CU, or 9 per EU. However, filling as much as possible
the remaining 40-36=4 wave slots without decreasing the number of groups
reveals that a larger group of 640 items yields 40 waves on the CU, or
10 per EU.
- Similarly, a group of 1024 items requires 16 waves per group. Only 2
groups made up of 16 waves each can fit fully on a CU ay any given time,
for a total of 32 waves on the CU, or 8 per EU. However, removing as
many waves as possible from the groups without being able to fit another
equal-sized group on the CU reveals that a smaller group of 896 items
yields 28 waves on the CU, or 7 per EU.
Therefore the achievable occupancy range for this function is not [8,9]
as the group size bounds directly yield, but [7,10].
Naturally this change causes a lot of test churn as instruction
scheduling is driven by achievable occupancy estimates. In most unit
tests the flat workgroup size range is the default [1,1024] which,
ignoring potential LDS limitations, would previously produce a scalar
occupancy of 8 (derived from 1024) on a lot of targets, whereas we now
consider the maximum occupancy to be 10 in such cases. Most tests are
updated automatically and checked manually for sanity. I also manually
changed some non-automatically generated assertions when necessary.
Fixes#118220.
OPSEL ASM Syntax for v_cvt_scalef32_pk_f32_fp4 : opsel:[x,y,z]
where, x & y i.e. OPSEL[1 : 0] selects which src_byte to read.
OPSEL ASM Syntax for v_cvt_scalef32_pk_fp4_f32 : opsel:[a,b,c,d]
where, c & d i.e. OPSEL[3 : 2] selects which dst_byte to write.
Co-authored-by: Pravin Jagtap <Pravin.Jagtap@amd.com>
Change the names of the TableGen features to match the names used by
AMDGPUSubtarget. "Addressable" refers to the amount that can be accessed
by a single workgroup. Add some explanatory comments. NFC.
A new function attribute named amdgpu_num_work_groups is added. This
attribute, which consists of three integers, allows programmers to let
the compiler know the number of workgroups to be launched in each of the
three dimensions and do optimizations based on that information.
---------
Co-authored-by: Jun Wang <jun.wang7@amd.com>
Real True16 instructions are as they are defined in the ISA. Fake True16
instructions are identical to real ones except that they take 32-bit
registers as operands and always use their low halves.
Reviewed By: Joe_Nash
Differential Revision: https://reviews.llvm.org/D156100
Previously we expanded these in a fast-math way and the device
libraries were relying on this behavior. The libraries have a pending
change to switch to the new target intrinsic.
Unlike the library version, this takes advantage of no-infinities on
the result overflow check.
This will do a value range merging down the callgraph, unlike the
current pass which can only propagate values to undecorated functions
from a kernel.
This one is a bit weird due to the interaction with the implied range
from amdgpu-flat-workgroup-size. At the default group range of 1,1024,
the minimum implied bounds is 4 so this ends up introducing the
attribute on undecorated functions. We could probably simplify this by
ignoring it and propagating the raw values. The subtarget interaction
and the interaction with amdgpu-flat-workgroup-size only really clamp
invalid values (plus the lower bound doesn't seem to do anything as
far as I can tell anyway).
Occupancy is expressed as waves per SIMD. This means that we need to
take into account the number of SIMDs per "CU" or, to be more precise,
the number of SIMDs over which a workgroup may be distributed.
getOccupancyWithLocalMemSize was wrong because it didn't take SIMDs
into account at all.
At the same time, we need to take into account that WGP mode offers
access to a larger total amount of LDS, since this can affect how
non-power-of-two LDS allocations are rounded. To make this work
consistently, we distinguish between (available) local memory size and
addressable local memory size (which is always limited by 64kB on
gfx10+, even with WGP mode).
This change results in a massive amount of test churn. A lot of it is
caused by the fact that the default work group size is 1024, which means
that (due to rounding effects) the default occupancy on older hardware
is 8 instead of 10, which affects scheduling via register pressure
estimates. I've adjusted most tests by just running the UTC tools, but
in some cases I manually changed the work group size to 32 or 64 to make
sure that work group size chunkiness has no effect.
Differential Revision: https://reviews.llvm.org/D139468
We will use this for more accurate occupancy computations. Note that
IsaInfo takes WGP mode vs. CU mode into account on gfx10+.
Differential Revision: https://reviews.llvm.org/D139467
Tablegen definitions for subtarget features and cpp predicate functions to
access the features.
New Sub-TargetProcessors and common latencies.
Simple changes to MIR codegen tests which pass on gfx11 because they have the
same output as previous subtargets or operate on pseudo instructions which
are reused from previous subtargets.
Contributors:
Jay Foad <jay.foad@amd.com>
Petar Avramovic <Petar.Avramovic@amd.com>
Patch 4/N for upstreaming of AMDGPU gfx11 architecture
Depends on D124538
Reviewed By: Petar.Avramovic, foad
Differential Revision: https://reviews.llvm.org/D125261
Use default member initializers in AMDGPUSubtarget and subclasses. This
is to guard against adding a new feature boolean in AMDGPUSubtarget.h
but forgetting to initialize it to false in AMDGPUSubtarget.cpp.
This was mostly autogenerated by:
clang-tidy -checks=-*,cppcoreguidelines-prefer-member-initializer,modernize-use-default-member-init -header-filter=Subtarget -fix lib/Target/AMDGPU/*Subtarget.cpp
Differential Revision: https://reviews.llvm.org/D123613
This was inheriting the mesa behavior, and as far as I know nobody is
using opencl kernels with amdpal. The isMesaKernel check was
irrelevant because this property needs to be held for all functions.
Add an overload to pass the flat workgroup range in separately. This
will allow the attributor to use the assumed value for
amdgpu-flat-workgroup-sizes when inferring amdgpu-waves-per-eu.
Added support to check if architecture supports s_mulhi which is used as part of
the decision whether or not to use valu 24 bit mul (if the mulhi gets
transformed to a valu op anyway, then may as well use it).
This is an extension of the work in D97063
Differential Revision: https://reviews.llvm.org/D103321
Change-Id: I80b1323de640a52623d69ac005a97d06a5d42a14
Currently, the compiler crashes in instruction selection of global
load/stores in gfx600 due to the lack of FLAT instructions. This patch
fix the crash by selecting MUBUF instructions for global load/stores
in gfx600.
Authored-by: Praveen Velliengiri <Praveen.Velliengiri@amd.com>
Reviewed by: t-tye
Differential revision: https://reviews.llvm.org/D92483
- Add an internal option `-amdgpu-use-aa-in-codegen` to enable or
disable this feature. By Default, it's enabled.
Differential Revision: https://reviews.llvm.org/D89320
The support is disabled by default. So far there is instruction
selection, spilling, and frame elimination. It also changes SP
from unswizzled to swizzled as used by flat scratch instructions,
so it cannot be mixed with MUBUF stack access.
At the very least missing:
- GlobalISel;
- Some optimizations in frame elimination in between vector
and scalar ALU;
- It shall finally allow to always materialize frame index
as an SGPR, but that is not implemented and frame elimination
cannot handle it yet;
- Unaligned and/or multidword flat scratch shall work, but it
is legalized now for MUBUF;
- Operand folding cannot optimize FI like with MUBUF yet;
- It will need scaling the value of the SP/FP in the DWARF
expression to recover the unswizzled scratch address;
Differential Revision: https://reviews.llvm.org/D89170
GFX10 enables third addressing mode for flat scratch instructions,
an ST mode. In that mode both register operands are omitted and
only swizzled offset is used in addition to flat_scratch base.
Differential Revision: https://reviews.llvm.org/D89501
Summary:
This implements a workaround for a hardware bug in gfx8 and gfx9,
where register usage is not estimated correctly for image_store and
image_gather4 instructions when D16 is used.
Change-Id: I4e30744da6796acac53a9b5ad37ac1c2035c8899
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D81172
