llvm-project.git/llvm/test/Transforms/SLPVectorizer/AMDGPU/reduction.ll, branch users/nico/python-2

[SLP]Improved reduction cost/codegen

2025-02-14T19:03:33+00:00

SLP vectorizer is able to combine several reductions from the list of
(potentially) reduced values with the different opcodes/values kind.
Currently, these reductions are handled independently of each other. But
instead the compiler can combine them into wide vector operations and
then perform only single reduction.
E.g, if the SLP vectorizer emits currently something like:
```
%r1 = reduce.add(<4 x i32> %v1)
%r2 = reduce.add(<4 x i32> %v2)
%r = add i32 %r1, %r2
```

it can be emitted as:
```
%v = add <4 x i32> %v1, %v2
%r = reduce.add(<4 x i32> %v)
```

It allows to improve the performance in some cases.

AVX512, -O3+LTO
Metric: size..text

Program                                                                                           size..text
                                                                                                  results     results0    diff
                      test-suite :: SingleSource/Benchmarks/Shootout-C++/Shootout-C++-matrix.test     4553.00     4615.00  1.4%
                                 test-suite :: SingleSource/Benchmarks/Adobe-C++/loop_unroll.test   412708.00   416820.00  1.0%
        test-suite :: SingleSource/UnitTests/Vector/AVX512BWVL/Vector-AVX512BWVL-mask_set_bw.test    12901.00    12981.00  0.6%
                        test-suite :: MultiSource/Benchmarks/FreeBench/fourinarow/fourinarow.test    22717.00    22813.00  0.4%
                             test-suite :: MultiSource/Benchmarks/mediabench/gsm/toast/toast.test    39722.00    39850.00  0.3%
                      test-suite :: MultiSource/Benchmarks/MiBench/telecomm-gsm/telecomm-gsm.test    39725.00    39853.00  0.3%
test-suite :: SingleSource/Regression/C/gcc-c-torture/execute/GCC-C-execute-builtin-bitops-1.test    15918.00    15967.00  0.3%
                                       test-suite :: External/SPEC/CFP2006/433.milc/433.milc.test   155491.00   155587.00  0.1%
                                     test-suite :: MicroBenchmarks/ImageProcessing/Blur/blur.test   227894.00   227942.00  0.0%
                                    test-suite :: MultiSource/Benchmarks/7zip/7zip-benchmark.test  1062188.00  1062364.00  0.0%
                                test-suite :: External/SPEC/CINT2006/464.h264ref/464.h264ref.test   793672.00   793720.00  0.0%
                              test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test   657371.00   657403.00  0.0%
                             test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test   657371.00   657403.00  0.0%
                   test-suite :: External/SPEC/CINT2017speed/600.perlbench_s/600.perlbench_s.test  2074917.00  2074933.00  0.0%
                    test-suite :: External/SPEC/CINT2017rate/500.perlbench_r/500.perlbench_r.test  2074917.00  2074933.00  0.0%
                                     test-suite :: MultiSource/Applications/JM/lencod/lencod.test   855219.00   855203.00 -0.0%

Benchmarks/Shootout-C++ - same transformed reduction
Adobe-C++/loop_unroll - same transformed reductions, new vector code
AVX512BWVL/Vector-AVX512BWVL-mask_set_bw - same transformed reductions
FreeBench/fourinarow - same transformed reductions
MiBench/telecomm-gsm - same transformed reductions
execute/GCC-C-execute-builtin-bitops-1 - same transformed reductions
CFP2006/433.milc - better vector code, several x i64 reductions + trunc
to i32 gets trunced to x i32 reductions
ImageProcessing/Blur - same transformed reductions
Benchmarks/7zip - same transformed reductions, extra 4 x vectorization
CINT2006/464.h264ref - same transformed reductions
CINT2017rate/525.x264_r
CINT2017speed/625.x264_s - same transformed reductions
CINT2017speed/600.perlbench_s
CINT2017rate/500.perlbench_r - transformed same reduction
JM/lencod - extra 4 x vectorization

RISC-V, SiFive-p670, -O3+LTO

Metric: size..text

Program                                                                                           size..text
                                                                                                  results    results0   diff
test-suite :: SingleSource/Regression/C/gcc-c-torture/execute/GCC-C-execute-builtin-bitops-1.test    8990.00    9514.00   5.8%
                                test-suite :: External/SPEC/CINT2006/464.h264ref/464.h264ref.test  588504.00  588488.00  -0.0%
                    test-suite :: MultiSource/Benchmarks/MiBench/consumer-lame/consumer-lame.test  147464.00  147440.00  -0.0%
              test-suite :: MultiSource/Benchmarks/MiBench/automotive-susan/automotive-susan.test   21496.00   21492.00  -0.0%
                                     test-suite :: MicroBenchmarks/ImageProcessing/Blur/blur.test  165420.00  165372.00  -0.0%
                                    test-suite :: MultiSource/Benchmarks/7zip/7zip-benchmark.test  843928.00  843648.00  -0.0%
                                    test-suite :: External/SPEC/CINT2006/458.sjeng/458.sjeng.test  100712.00  100672.00  -0.0%
                      test-suite :: MultiSource/Benchmarks/MiBench/telecomm-gsm/telecomm-gsm.test   24384.00   24336.00  -0.2%
                             test-suite :: MultiSource/Benchmarks/mediabench/gsm/toast/toast.test   24380.00   24332.00  -0.2%
             test-suite :: SingleSource/UnitTests/Vectorizer/VPlanNativePath/outer-loop-vect.test   10348.00   10316.00  -0.3%
                                 test-suite :: SingleSource/Benchmarks/Adobe-C++/loop_unroll.test  221304.00  220480.00  -0.4%
                      test-suite :: SingleSource/Benchmarks/Shootout-C++/Shootout-C++-matrix.test    3750.00    3736.00  -0.4%
                            test-suite :: SingleSource/Regression/C/Regression-C-DuffsDevice.test     678.00     370.00 -45.4%

execute/GCC-C-execute-builtin-bitops-1 - extra 4 x reductions, same
transformed reductions
CINT2006/464.h264ref - extra 4 x reductions, same transformed reductions
MiBench/consumer-lame - 2 4 x i1 merged to 8 x i1 reductions (bitcast + ctpop)
MiBench/automotive-susan - same transformed reductions
ImageProcessing/Blur - same transformed reductions
Benchmarks/7zip - same transformed reductions
CINT2006/458.sjeng - 2 4 x i1 merged to 8 x i1 reductions (bitcast + ctpop)
MiBench/telecomm-gsm - same transformed reductions
Benchmarks/mediabench - same transformed reductions
Vectorizer/VPlanNativePath - same transformed reductions
Adobe-C++/loop_unroll - extra 4 x reductions, same transformed reductions
Benchmarks/Shootout-C++ - extra 4 x reductions, same transformed reductions
Regression/C/Regression-C-DuffsDevice - same transformed reductions

Reviewers: hiraditya, topperc, preames

Pull Request: https://github.com/llvm/llvm-project/pull/118293

Revert "[SLP]Improved reduction cost/codegen"

2025-02-14T18:18:07+00:00

This reverts commit 7ec60bf0166519317b5ae2505dd6ed4660e3ea39 to fix
a bug reported in https://github.com/llvm/llvm-project/issues/127220.

[SLP]Improved reduction cost/codegen

2025-02-14T13:15:29+00:00

SLP vectorizer is able to combine several reductions from the list of
(potentially) reduced values with the different opcodes/values kind.
Currently, these reductions are handled independently of each other. But
instead the compiler can combine them into wide vector operations and
then perform only single reduction.
E.g, if the SLP vectorizer emits currently something like:
```
%r1 = reduce.add(<4 x i32> %v1)
%r2 = reduce.add(<4 x i32> %v2)
%r = add i32 %r1, %r2
```

it can be emitted as:
```
%v = add <4 x i32> %v1, %v2
%r = reduce.add(<4 x i32> %v)
```

It allows to improve the performance in some cases.

AVX512, -O3+LTO
Metric: size..text

Program                                                                                           size..text
                                                                                                  results     results0    diff
                      test-suite :: SingleSource/Benchmarks/Shootout-C++/Shootout-C++-matrix.test     4553.00     4615.00  1.4%
                                 test-suite :: SingleSource/Benchmarks/Adobe-C++/loop_unroll.test   412708.00   416820.00  1.0%
        test-suite :: SingleSource/UnitTests/Vector/AVX512BWVL/Vector-AVX512BWVL-mask_set_bw.test    12901.00    12981.00  0.6%
                        test-suite :: MultiSource/Benchmarks/FreeBench/fourinarow/fourinarow.test    22717.00    22813.00  0.4%
                             test-suite :: MultiSource/Benchmarks/mediabench/gsm/toast/toast.test    39722.00    39850.00  0.3%
                      test-suite :: MultiSource/Benchmarks/MiBench/telecomm-gsm/telecomm-gsm.test    39725.00    39853.00  0.3%
test-suite :: SingleSource/Regression/C/gcc-c-torture/execute/GCC-C-execute-builtin-bitops-1.test    15918.00    15967.00  0.3%
                                       test-suite :: External/SPEC/CFP2006/433.milc/433.milc.test   155491.00   155587.00  0.1%
                                     test-suite :: MicroBenchmarks/ImageProcessing/Blur/blur.test   227894.00   227942.00  0.0%
                                    test-suite :: MultiSource/Benchmarks/7zip/7zip-benchmark.test  1062188.00  1062364.00  0.0%
                                test-suite :: External/SPEC/CINT2006/464.h264ref/464.h264ref.test   793672.00   793720.00  0.0%
                              test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test   657371.00   657403.00  0.0%
                             test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test   657371.00   657403.00  0.0%
                   test-suite :: External/SPEC/CINT2017speed/600.perlbench_s/600.perlbench_s.test  2074917.00  2074933.00  0.0%
                    test-suite :: External/SPEC/CINT2017rate/500.perlbench_r/500.perlbench_r.test  2074917.00  2074933.00  0.0%
                                     test-suite :: MultiSource/Applications/JM/lencod/lencod.test   855219.00   855203.00 -0.0%

Benchmarks/Shootout-C++ - same transformed reduction
Adobe-C++/loop_unroll - same transformed reductions, new vector code
AVX512BWVL/Vector-AVX512BWVL-mask_set_bw - same transformed reductions
FreeBench/fourinarow - same transformed reductions
MiBench/telecomm-gsm - same transformed reductions
execute/GCC-C-execute-builtin-bitops-1 - same transformed reductions
CFP2006/433.milc - better vector code, several x i64 reductions + trunc
to i32 gets trunced to x i32 reductions
ImageProcessing/Blur - same transformed reductions
Benchmarks/7zip - same transformed reductions, extra 4 x vectorization
CINT2006/464.h264ref - same transformed reductions
CINT2017rate/525.x264_r
CINT2017speed/625.x264_s - same transformed reductions
CINT2017speed/600.perlbench_s
CINT2017rate/500.perlbench_r - transformed same reduction
JM/lencod - extra 4 x vectorization

RISC-V, SiFive-p670, -O3+LTO

Metric: size..text

Program                                                                                           size..text
                                                                                                  results    results0   diff
test-suite :: SingleSource/Regression/C/gcc-c-torture/execute/GCC-C-execute-builtin-bitops-1.test    8990.00    9514.00   5.8%
                                test-suite :: External/SPEC/CINT2006/464.h264ref/464.h264ref.test  588504.00  588488.00  -0.0%
                    test-suite :: MultiSource/Benchmarks/MiBench/consumer-lame/consumer-lame.test  147464.00  147440.00  -0.0%
              test-suite :: MultiSource/Benchmarks/MiBench/automotive-susan/automotive-susan.test   21496.00   21492.00  -0.0%
                                     test-suite :: MicroBenchmarks/ImageProcessing/Blur/blur.test  165420.00  165372.00  -0.0%
                                    test-suite :: MultiSource/Benchmarks/7zip/7zip-benchmark.test  843928.00  843648.00  -0.0%
                                    test-suite :: External/SPEC/CINT2006/458.sjeng/458.sjeng.test  100712.00  100672.00  -0.0%
                      test-suite :: MultiSource/Benchmarks/MiBench/telecomm-gsm/telecomm-gsm.test   24384.00   24336.00  -0.2%
                             test-suite :: MultiSource/Benchmarks/mediabench/gsm/toast/toast.test   24380.00   24332.00  -0.2%
             test-suite :: SingleSource/UnitTests/Vectorizer/VPlanNativePath/outer-loop-vect.test   10348.00   10316.00  -0.3%
                                 test-suite :: SingleSource/Benchmarks/Adobe-C++/loop_unroll.test  221304.00  220480.00  -0.4%
                      test-suite :: SingleSource/Benchmarks/Shootout-C++/Shootout-C++-matrix.test    3750.00    3736.00  -0.4%
                            test-suite :: SingleSource/Regression/C/Regression-C-DuffsDevice.test     678.00     370.00 -45.4%

execute/GCC-C-execute-builtin-bitops-1 - extra 4 x reductions, same
transformed reductions
CINT2006/464.h264ref - extra 4 x reductions, same transformed reductions
MiBench/consumer-lame - 2 4 x i1 merged to 8 x i1 reductions (bitcast + ctpop)
MiBench/automotive-susan - same transformed reductions
ImageProcessing/Blur - same transformed reductions
Benchmarks/7zip - same transformed reductions
CINT2006/458.sjeng - 2 4 x i1 merged to 8 x i1 reductions (bitcast + ctpop)
MiBench/telecomm-gsm - same transformed reductions
Benchmarks/mediabench - same transformed reductions
Vectorizer/VPlanNativePath - same transformed reductions
Adobe-C++/loop_unroll - extra 4 x reductions, same transformed reductions
Benchmarks/Shootout-C++ - extra 4 x reductions, same transformed reductions
Regression/C/Regression-C-DuffsDevice - same transformed reductions

Reviewers: hiraditya, topperc, preames

Pull Request: https://github.com/llvm/llvm-project/pull/118293

Revert "[SLP]Improved reduction cost/codegen"

2025-02-14T12:12:47+00:00

This reverts commit 2ad816648f2719e6c0da507a1a371f2cad4a3f1c to fix
bug/miscompiles, reported in
https://github.com/llvm/llvm-project/pull/118293#issuecomment-2658906033
and https://github.com/llvm/llvm-project/pull/118293#issuecomment-2659024785.

[SLP]Improved reduction cost/codegen

2025-02-13T15:36:28+00:00

SLP vectorizer is able to combine several reductions from the list of
(potentially) reduced values with the different opcodes/values kind.
Currently, these reductions are handled independently of each other. But
instead the compiler can combine them into wide vector operations and
then perform only single reduction.
E.g, if the SLP vectorizer emits currently something like:
```
%r1 = reduce.add(<4 x i32> %v1)
%r2 = reduce.add(<4 x i32> %v2)
%r = add i32 %r1, %r2
```

it can be emitted as:
```
%v = add <4 x i32> %v1, %v2
%r = reduce.add(<4 x i32> %v)
```

It allows to improve the performance in some cases.

AVX512, -O3+LTO
Metric: size..text

Program                                                                                           size..text
                                                                                                  results     results0    diff
                      test-suite :: SingleSource/Benchmarks/Shootout-C++/Shootout-C++-matrix.test     4553.00     4615.00  1.4%
                                 test-suite :: SingleSource/Benchmarks/Adobe-C++/loop_unroll.test   412708.00   416820.00  1.0%
        test-suite :: SingleSource/UnitTests/Vector/AVX512BWVL/Vector-AVX512BWVL-mask_set_bw.test    12901.00    12981.00  0.6%
                        test-suite :: MultiSource/Benchmarks/FreeBench/fourinarow/fourinarow.test    22717.00    22813.00  0.4%
                             test-suite :: MultiSource/Benchmarks/mediabench/gsm/toast/toast.test    39722.00    39850.00  0.3%
                      test-suite :: MultiSource/Benchmarks/MiBench/telecomm-gsm/telecomm-gsm.test    39725.00    39853.00  0.3%
test-suite :: SingleSource/Regression/C/gcc-c-torture/execute/GCC-C-execute-builtin-bitops-1.test    15918.00    15967.00  0.3%
                                       test-suite :: External/SPEC/CFP2006/433.milc/433.milc.test   155491.00   155587.00  0.1%
                                     test-suite :: MicroBenchmarks/ImageProcessing/Blur/blur.test   227894.00   227942.00  0.0%
                                    test-suite :: MultiSource/Benchmarks/7zip/7zip-benchmark.test  1062188.00  1062364.00  0.0%
                                test-suite :: External/SPEC/CINT2006/464.h264ref/464.h264ref.test   793672.00   793720.00  0.0%
                              test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test   657371.00   657403.00  0.0%
                             test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test   657371.00   657403.00  0.0%
                   test-suite :: External/SPEC/CINT2017speed/600.perlbench_s/600.perlbench_s.test  2074917.00  2074933.00  0.0%
                    test-suite :: External/SPEC/CINT2017rate/500.perlbench_r/500.perlbench_r.test  2074917.00  2074933.00  0.0%
                                     test-suite :: MultiSource/Applications/JM/lencod/lencod.test   855219.00   855203.00 -0.0%

Benchmarks/Shootout-C++ - same transformed reduction
Adobe-C++/loop_unroll - same transformed reductions, new vector code
AVX512BWVL/Vector-AVX512BWVL-mask_set_bw - same transformed reductions
FreeBench/fourinarow - same transformed reductions
MiBench/telecomm-gsm - same transformed reductions
execute/GCC-C-execute-builtin-bitops-1 - same transformed reductions
CFP2006/433.milc - better vector code, several x i64 reductions + trunc
to i32 gets trunced to x i32 reductions
ImageProcessing/Blur - same transformed reductions
Benchmarks/7zip - same transformed reductions, extra 4 x vectorization
CINT2006/464.h264ref - same transformed reductions
CINT2017rate/525.x264_r
CINT2017speed/625.x264_s - same transformed reductions
CINT2017speed/600.perlbench_s
CINT2017rate/500.perlbench_r - transformed same reduction
JM/lencod - extra 4 x vectorization

RISC-V, SiFive-p670, -O3+LTO

Metric: size..text

Program                                                                                           size..text
                                                                                                  results    results0   diff
test-suite :: SingleSource/Regression/C/gcc-c-torture/execute/GCC-C-execute-builtin-bitops-1.test    8990.00    9514.00   5.8%
                                test-suite :: External/SPEC/CINT2006/464.h264ref/464.h264ref.test  588504.00  588488.00  -0.0%
                    test-suite :: MultiSource/Benchmarks/MiBench/consumer-lame/consumer-lame.test  147464.00  147440.00  -0.0%
              test-suite :: MultiSource/Benchmarks/MiBench/automotive-susan/automotive-susan.test   21496.00   21492.00  -0.0%
                                     test-suite :: MicroBenchmarks/ImageProcessing/Blur/blur.test  165420.00  165372.00  -0.0%
                                    test-suite :: MultiSource/Benchmarks/7zip/7zip-benchmark.test  843928.00  843648.00  -0.0%
                                    test-suite :: External/SPEC/CINT2006/458.sjeng/458.sjeng.test  100712.00  100672.00  -0.0%
                      test-suite :: MultiSource/Benchmarks/MiBench/telecomm-gsm/telecomm-gsm.test   24384.00   24336.00  -0.2%
                             test-suite :: MultiSource/Benchmarks/mediabench/gsm/toast/toast.test   24380.00   24332.00  -0.2%
             test-suite :: SingleSource/UnitTests/Vectorizer/VPlanNativePath/outer-loop-vect.test   10348.00   10316.00  -0.3%
                                 test-suite :: SingleSource/Benchmarks/Adobe-C++/loop_unroll.test  221304.00  220480.00  -0.4%
                      test-suite :: SingleSource/Benchmarks/Shootout-C++/Shootout-C++-matrix.test    3750.00    3736.00  -0.4%
                            test-suite :: SingleSource/Regression/C/Regression-C-DuffsDevice.test     678.00     370.00 -45.4%

execute/GCC-C-execute-builtin-bitops-1 - extra 4 x reductions, same
transformed reductions
CINT2006/464.h264ref - extra 4 x reductions, same transformed reductions
MiBench/consumer-lame - 2 4 x i1 merged to 8 x i1 reductions (bitcast + ctpop)
MiBench/automotive-susan - same transformed reductions
ImageProcessing/Blur - same transformed reductions
Benchmarks/7zip - same transformed reductions
CINT2006/458.sjeng - 2 4 x i1 merged to 8 x i1 reductions (bitcast + ctpop)
MiBench/telecomm-gsm - same transformed reductions
Benchmarks/mediabench - same transformed reductions
Vectorizer/VPlanNativePath - same transformed reductions
Adobe-C++/loop_unroll - extra 4 x reductions, same transformed reductions
Benchmarks/Shootout-C++ - extra 4 x reductions, same transformed reductions
Regression/C/Regression-C-DuffsDevice - same transformed reductions

Reviewers: hiraditya, topperc, preames

Pull Request: https://github.com/llvm/llvm-project/pull/118293

Prefer use of 0.0 over -0.0 for fadd reductions w/nsz (in IR) (#106770)

2024-09-03T16:16:37+00:00

This is a follow up to 924907bc6, and is mostly motivated by consistency
but does include one additional optimization. In general, we prefer 0.0
over -0.0 as the identity value for an fadd. We use that value in
several places, but don't in others. So, let's be consistent and use the
same identity (when nsz allows) everywhere.

This creates a bunch of test churn, but due to 924907bc6, most of that
churn doesn't actually indicate a change in codegen. The exception is
that this change enables the use of 0.0 for nsz, but *not* reasoc, fadd
reductions. Or said differently, it allows the neutral value of an
ordered fadd reduction to be 0.0.

[SLP]Improve/fix extracts calculations for non-power-of-2 elements.

2024-05-24T13:33:36+00:00

One of the previous patches introduced initial support for non-power-of-2
number of elements but some parts of the SLP vectorizer still were not
adjusted to handle the costs correctly. Patch fixes it by improving
analysis of the non-power-of-2 number of elements and fixes in the cost
of the extractelements instructions.

Reviewers: RKSimon

Reviewed By: RKSimon

Pull Request: https://github.com/llvm/llvm-project/pull/93213

[AMDGPU] Vectorize more 16 bit shuffles (#90648)

2024-05-21T16:21:36+00:00

In the case of larger vectors, we should still prefer the vectorized
version (i.e. shufflevector vs extract/insert chains).

In arithmetic chains, vectorization results in chains of packed math
instructions (as opposed to unpack/repack & scalarized arithmetic):
https://godbolt.org/z/c5onaf6G5

In chains with PHIs, vectorization again removes the unnecessary pack /
repack code around BBs: https://godbolt.org/z/vz7zYzvhs

[NFC] Port all SLPVectorizer tests to `-passes=` syntax

2022-12-07T18:44:09+00:00

[Vectorizers][TTI] remove option to bypass creation of vector reduction intrinsics

2021-02-12T13:13:50+00:00

The vector reduction intrinsics started life as experimental ops, so backend support
was lacking. As part of promoting them to 1st-class intrinsics, however, codegen
support was added/improved:
D58015
D90247

So I think it is safe to now remove this complication from IR.

Note that we still have an IR-level codegen expansion pass for these as discussed
in D95690. Removing that is another step in simplifying the logic. Also note that
x86 was already unconditionally forming reductions in IR, so there should be no
difference for x86.

I spot checked a couple of the tests here by running them through opt+llc and did
not see any asm diffs.

If we do find functional differences for other targets, it should be possible
to (at least temporarily) restore the shuffle IR with the ExpandReductions IR
pass.

Differential Revision: https://reviews.llvm.org/D96552