From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-bugzilla@gcc.gnu.org>
Received: by sourceware.org (Postfix, from userid 48)
	id EE16A384641E; Thu, 25 Apr 2024 12:28:42 +0000 (GMT)
DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org EE16A384641E
DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gcc.gnu.org;
	s=default; t=1714048122;
	bh=w2QjRIuxaMyF/X/sRVnZN+dPu1Ue4Y9TZW8e7X18zNk=;
	h=From:To:Subject:Date:From;
	b=yrayweQdCPAF98a6AbLrT6TbAR2fhu9Xo9U53HKqQlfRVUh2W3ln2vHA22/iltAXo
	 7l4MglWYQkupcSxQ5ZkEgAaCsw/V8JFOud19b+6STYwPAEe3hcCWo40OUYWslaN9FY
	 8C+Tbl+ncEmsweTqbwQ7D+lMVO8mBPpEYEt99Arc=
From: "hubicka at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>
To: gcc-bugs@gcc.gnu.org
Subject: [Bug middle-end/114852] New: jpegxl 10.0.1 is faster with clang18
 then with gcc14
Date: Thu, 25 Apr 2024 12:28:42 +0000
X-Bugzilla-Reason: CC
X-Bugzilla-Type: new
X-Bugzilla-Watch-Reason: None
X-Bugzilla-Product: gcc
X-Bugzilla-Component: middle-end
X-Bugzilla-Version: 14.0
X-Bugzilla-Keywords: 
X-Bugzilla-Severity: normal
X-Bugzilla-Who: hubicka at gcc dot gnu.org
X-Bugzilla-Status: UNCONFIRMED
X-Bugzilla-Resolution: 
X-Bugzilla-Priority: P3
X-Bugzilla-Assigned-To: unassigned at gcc dot gnu.org
X-Bugzilla-Target-Milestone: ---
X-Bugzilla-Flags: 
X-Bugzilla-Changed-Fields: bug_id short_desc product version bug_status
 bug_severity priority component assigned_to reporter target_milestone
Message-ID: <bug-114852-4@http.gcc.gnu.org/bugzilla/>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
X-Bugzilla-URL: http://gcc.gnu.org/bugzilla/
Auto-Submitted: auto-generated
MIME-Version: 1.0
List-Id: <gcc-bugs.sourceware.org>

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=3D114852

            Bug ID: 114852
           Summary: jpegxl 10.0.1 is faster with clang18 then with gcc14
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hubicka at gcc dot gnu.org
  Target Milestone: ---

https://www.phoronix.com/review/gcc14-clang18-amd-zen4/3
reports about 8% difference.  I can measure 13% on zen3.  The code has chan=
ged
and it is no longer bound by push_back but runs AVX2 version of inner loops.

The hottest loops looks comparable

  0.00 =E2=94=82266:=E2=94=8C=E2=94=80=E2=86=92vmovaps      (%r14,%rax,4),%=
ymm0
  0.11 =E2=94=82    =E2=94=82  vmulps       (%rcx,%rax,4),%ymm7,%ymm2
  1.18 =E2=94=82    =E2=94=82  vfnmadd213ps (%rsi,%rax,4),%ymm11,%ymm0
  0.25 =E2=94=82    =E2=94=82  vmulps       %ymm2,%ymm0,%ymm0
  5.94 =E2=94=82    =E2=94=82  vroundps     $0x8,%ymm0,%ymm2
  0.35 =E2=94=82    =E2=94=82  vsubps       %ymm2,%ymm0,%ymm0
  1.05 =E2=94=82    =E2=94=82  vmulps       (%rdx,%rax,4),%ymm0,%ymm0
  3.19 =E2=94=82    =E2=94=82  vmovaps      %ymm0,0x0(%r13,%rax,4)
  0.15 =E2=94=82    =E2=94=82  vandps       %ymm10,%ymm2,%ymm0
  0.03 =E2=94=82    =E2=94=82  add          $0x8,%rax
  0.03 =E2=94=82    =E2=94=82  vcmpeqps     %ymm8,%ymm0,%ymm2
  0.09 =E2=94=82    =E2=94=82  vsqrtps      %ymm0,%ymm0
 27.25 =E2=94=82    =E2=94=82  vaddps       %ymm0,%ymm6,%ymm6
  0.35 =E2=94=82    =E2=94=82  vandnps      %ymm9,%ymm2,%ymm0
  0.12 =E2=94=82    =E2=94=82  vaddps       %ymm0,%ymm5,%ymm5
  0.05 =E2=94=82    =E2=94=9C=E2=94=80=E2=94=80cmp          %r12,%rax
  0.02 =E2=94=82    =E2=94=94=E2=94=80=E2=94=80jb           266

and clang

  0.00 =E2=94=82 c90:=E2=94=8C=E2=94=80=E2=86=92vmulps       (%r9,%rdx,4),%=
ymm0,%ymm2
  0.97 =E2=94=82     =E2=94=82  vmovaps      (%r15,%rdx,4),%ymm1
  0.36 =E2=94=82     =E2=94=82  vsubps       %ymm2,%ymm1,%ymm1
  4.24 =E2=94=82     =E2=94=82  vmulps       (%rcx,%rdx,4),%ymm4,%ymm2
  1.92 =E2=94=82     =E2=94=82  vmulps       %ymm2,%ymm1,%ymm1
  0.65 =E2=94=82     =E2=94=82  vroundps     $0x8,%ymm1,%ymm2
  0.06 =E2=94=82     =E2=94=82  vsubps       %ymm2,%ymm1,%ymm1
  1.11 =E2=94=82     =E2=94=82  vmulps       (%rax,%rdx,4),%ymm1,%ymm1
  3.53 =E2=94=82     =E2=94=82  vmovaps      %ymm1,(%rsi,%rdx,4)
  0.68 =E2=94=82     =E2=94=82  vandps       %ymm6,%ymm2,%ymm1
  0.23 =E2=94=82     =E2=94=82  vcmpneqps    %ymm5,%ymm2,%ymm2
  3.64 =E2=94=82     =E2=94=82  add          $0x8,%rdx
  0.24 =E2=94=82     =E2=94=82  vsqrtps      %ymm1,%ymm1
 22.16 =E2=94=82     =E2=94=82  vaddps       %ymm1,%ymm8,%ymm8
  0.25 =E2=94=82     =E2=94=82  vbroadcastss 0x31eba5(%rip),%ymm1        # =
34f840
<jxl::palette_internal::GetPaletteValue(int const*, int, unsigned long, int,
int, int)::kMultiplier+0xe0>
  0.05 =E2=94=82     =E2=94=82  vandps       %ymm1,%ymm2,%ymm1
  0.04 =E2=94=82     =E2=94=82  vaddps       %ymm1,%ymm7,%ymm7
  0.11 =E2=94=82     =E2=94=9C=E2=94=80=E2=94=80cmp          %rdi,%rdx
  0.07 =E2=94=82     =E2=94=94=E2=94=80=E2=94=80jb           c90=E2=96=92

GCC profile:
  10.78%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::EstimateEntropy(jxl::AcStrategy const&, float, unsigned long,
unsigned long, jxl::ACSConfig const&, float con
   7.02%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::FindBestMultiplier(float const*, float const*, unsigned long,
float, float, bool) [clone .part.0]
   4.50%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::Symmetric5Row<jxl::WrapUnchanged>(jxl::Plane<float> const&,
jxl::RectT<unsigned long> const&, long, jxl:
   4.47%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::(anonymous namespace)::TransformFromPixels(jxl::AcStrategy::Ty=
pe,
float const*, unsigned long, float*, float*
   4.31%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::(anonymous namespace)::TransformToPixels(jxl::AcStrategy::Type,
float*, float*, unsigned long, float*)
   4.00%  cjxl             libjxl.so.0.10.1               [.]
jxl::ThreadPool::RunCallState<jxl::Status (unsigned long),
jxl::ConvertFromExternalNoSizeCheck(unsigned char const*, unsig
   3.64%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::TokenizeCoefficients(unsigned int const*, jxl::RectT<unsigned
long> const&, int const* restrict*, jxl::AcStra
   3.56%  cjxl             libm.so.6                      [.] __ieee754_pow=
_fma
   3.49%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::(anonymous namespace)::IDCT1DImpl<8ul, 8ul>::operator()(float
const*, unsigned long, float*, unsigned long, f
   3.43%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::(anonymous
namespace)::AdaptiveQuantizationImpl::ComputeTile(float, float,
jxl::Image3<float> const&, jxl::Re
   3.27%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::(anonymous namespace)::DCT1DWrapper<32ul, 0ul,
jxl::N_AVX2::(anonymous namespace)::DCTFrom, jxl::N_AVX2:
   3.16%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::(anonymous namespace)::DCT1DImpl<8ul, 8ul>::operator()(float*,
float*) [clone .isra.0]
   2.87%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::(anonymous namespace)::ComputeScaledIDCT<4ul,
8ul>::operator()<jxl::N_AVX2::(anonymous namespace)::DCTTo
   2.83%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::(anonymous namespace)::DCT1DWrapper<64ul, 0ul,
jxl::N_AVX2::(anonymous namespace)::DCTFrom, jxl::N_AVX2:
   2.22%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::AdjustQuantBlockAC(jxl::Quantizer const&, unsigned long, float,
unsigned long, unsigned long, unsigned long,
   2.17%  cjxl             libm.so.6                      [.] __log1pf
   2.02%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::(anonymous namespace)::IDCT1DWrapper<64ul, 0ul,
jxl::N_AVX2::(anonymous namespace)::DCTFrom, jxl::N_AVX2
   1.89%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::(anonymous namespace)::ComputeScaledDCT<4ul,
8ul>::operator()<jxl::N_AVX2::(anonymous namespace)::DCTFro
   1.84%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::(anonymous namespace)::DCT1DWrapper<8ul, 4ul,
jxl::N_AVX2::(anonymous namespace)::DCTFrom, jxl::N_AVX2::
   1.77%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::(anonymous namespace)::IDCT1DWrapper<32ul, 0ul,
jxl::N_AVX2::(anonymous namespace)::DCTFrom, jxl::N_AVX2
   1.74%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::(anonymous namespace)::ComputeScaledDCT<4ul,
4ul>::operator()<jxl::N_AVX2::(anonymous namespace)::DCTFro


clang profile:
  12.96%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::EstimateEntropy(jxl::AcStrategy const&, float, unsigned long,
unsigned long, jxl::ACSConfig const&, float const*, float*, float*, unsigned
int*)                                           =E2=97=86
   8.33%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::(anonymous namespace)::TransformToPixels(jxl::AcStrategy::Type,
float*, float*, unsigned long, float*)=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20
                                           =E2=96=92
   7.87%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::(anonymous namespace)::TransformFromPixels(jxl::AcStrategy::Ty=
pe,
float const*, unsigned long, float*, float*)=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20
                                         =E2=96=92
   7.78%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::ComputeTile(jxl::Image3<float> const&, jxl::RectT<unsigned lon=
g>
const&, jxl::DequantMatrices const&, jxl::AcStrategyImage const*,
jxl::Plane<int> const*, jxl::Quantizer const*, jxl::Rect=E2=96=92
   5.03%  cjxl             libjxl.so.0.10.1               [.]
jxl::ThreadPool::RunCallState<jxl::Status (unsigned long),
jxl::N_AVX2::Symmetric5(jxl::Plane<float> const&, jxl::RectT<unsigned long>
const&, jxl::WeightsSymmetric5 const&, jxl::ThreadPool*, jxl::Pla=E2=96=92
   4.66%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::(anonymous namespace)::DCT1DImpl<16ul, 8ul>::operator()(float*,
float*)=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20
                                           =E2=96=92
   4.56%  cjxl             libjxl.so.0.10.1               [.]
jxl::ThreadPool::RunCallState<jxl::Status (unsigned long),
jxl::ConvertFromExternalNoSizeCheck(unsigned char const*, unsigned long,
unsigned long, unsigned long, unsigned long, JxlPixelFormat, unsigne=E2=96=
=92
   4.23%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::(anonymous namespace)::IDCT1DImpl<16ul, 8ul>::operator()(float
const*, unsigned long, float*, unsigned long, float*)=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20
                                            =E2=96=92
   4.22%  cjxl             libm.so.6                      [.] __ieee754_pow=
_fma
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20
                      =E2=96=92
   3.62%  cjxl             libjxl.so.0.10.1               [.]
jxl::ThreadPool::RunCallState<jxl::N_AVX2::(anonymous
namespace)::AdaptiveQuantizationMap(float, jxl::Image3<float> const&,
jxl::RectT<unsigned long> const&, float, jxl::ThreadPool*, jxl::Plane<float=
>=E2=96=92
   3.04%  cjxl             libjxl.so.0.10.1               [.] void
jxl::N_AVX2::(anonymous namespace)::GenericTransposeBlock<0ul, 0ul,
jxl::N_AVX2::(anonymous namespace)::DCTFrom, jxl::N_AVX2::(anonymous
namespace)::DCTTo>(jxl::N_AVX2::(anonymous namespace)::Tra=E2=96=92
   2.84%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::AdjustQuantBlockAC(jxl::Quantizer const&, unsigned long, float,
unsigned long, unsigned long, unsigned long, float*, float const*, int*)=20=
=20=20=20=20=20=20
                                           =E2=96=92
   2.67%  cjxl             libjxl.so.0.10.1               [.]
jxl::ThreadPool::RunCallState<jxl::Status (unsigned long),
jxl::N_AVX2::SRGBToXYB(float const*, jxl::ThreadPool*,
jxl::Image3<float>*)::$_0>::CallDataFunc(void*, unsigned int, unsigned long=
)=20=20=20
      =E2=96=92
   2.51%  cjxl             libm.so.6                      [.] __log1pf=20=
=20=20=20=20=20=20=20=20
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20
                      =E2=96=92
   2.48%  cjxl             libjxl.so.0.10.1               [.]
jxl::N_AVX2::TokenizeCoefficients(unsigned int const*, jxl::RectT<unsigned
long> const&, int const* restrict*, jxl::AcStrategyImage const&,
jxl::YCbCrChromaSubsampling, jxl::Image3<int>*, std::vector<=E2=96=92=