Script 'mail_helper' called by obssrc Hello community, here is the log from the commit of package clpeak for openSUSE:Factory checked in at 2026-08-28 19:51:23 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Comparing /work/SRC/openSUSE:Factory/clpeak (Old) and /work/SRC/openSUSE:Factory/.clpeak.new.1265 (New) ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "clpeak" Fri Aug 28 19:51:23 2026 rev:20 rq:1374230 version:2.1.4 Changes: -------- --- /work/SRC/openSUSE:Factory/clpeak/clpeak.changes 2026-08-25 13:19:39.354041044 +0200 +++ /work/SRC/openSUSE:Factory/.clpeak.new.1265/clpeak.changes 2026-08-28 19:52:51.333455077 +0200 @@ -1,0 +2,25 @@ +Fri Aug 28 05:54:58 UTC 2026 - Martin Pluskal <[email protected]> + +- Update to 2.1.4, a measurement-accuracy release -- the corrected + rows below no longer compare with older runs: + * INT8 dot and mixed precision (all backends) each ran an op the + op budget never counted inside the timed chain: OpenCL mp read + half the fp32 rate, RTX 5060 CUDA int8 33587 -> 41683 GOPS. + * Image bandwidth races both walk orders and reports the faster; + the ~1.5x CUDA/Vulkan gap was deflation, not inflation. + * Vulkan coopmat: tiles run at the subgroup width the driver + advertised, saturating tiles are filtered out (fixes a + process-killing Arc A380 int8 crash), shaders target SPIR-V 1.5, + new fp16_f16acc row; A380 fp16 12.5 -> ~25.9 TFLOPS. + * Vulkan fp16 shaders get the fp32 register budget; affine chain + seeds step a full vector width, so width >= 2 chains stop + starting out bitwise identical and CSE-able. + * Global bandwidth sizes its working set to 8x the device LLC; + local bandwidth is skipped where there is no real scratchpad. + * OpenCL transfer bandwidth collapsed to h2d_pinned/d2h_pinned; + oneAPI INT8 dot dropped, SYCL has no DP4a intrinsic so it never + measured one. +- Drop clpeak-version-override.patch, merged upstream + (gh#krrishnarraj/clpeak#206); -DCLPEAK_VERSION_OVERRIDE stays. + +------------------------------------------------------------------- Old: ---- clpeak-2.1.3.tar.gz clpeak-version-override.patch New: ---- clpeak-2.1.4.tar.gz ----------(Old B)---------- Old: measured one. - Drop clpeak-version-override.patch, merged upstream (gh#krrishnarraj/clpeak#206); -DCLPEAK_VERSION_OVERRIDE stays. ----------(Old E)---------- ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Other differences: ------------------ ++++++ clpeak.spec ++++++ --- /var/tmp/diff_new_pack.AHJCRc/_old 2026-08-28 19:52:52.300488879 +0200 +++ /var/tmp/diff_new_pack.AHJCRc/_new 2026-08-28 19:52:52.302488949 +0200 @@ -18,7 +18,7 @@ Name: clpeak -Version: 2.1.3 +Version: 2.1.4 Release: 0 Summary: Find peak OpenCL capacities like bandwidth & compute # Legal-Review-Notice: upstream relicensed from the Unlicense to Apache-2.0 @@ -28,9 +28,6 @@ License: Apache-2.0 URL: https://github.com/krrishnarraj/clpeak Source: https://github.com/krrishnarraj/clpeak/archive/refs/tags/%{version}.tar.gz#/%{name}-%{version}.tar.gz -# PATCH-FIX-UPSTREAM clpeak-version-override.patch gh#krrishnarraj/clpeak#206 [email protected] -# -- let the builder pass the version in; tarball builds report "unknown" since 2.1.3 -Patch0: clpeak-version-override.patch BuildRequires: cmake >= 3.20 BuildRequires: gcc-c++ BuildRequires: ninja @@ -55,9 +52,9 @@ # oneAPI likewise auto-skip when their SDKs are absent — keep them off # explicitly. Vulkan + OpenCL + CPU stay on (defaults). %define __builder ninja -# GitHub archives carry no .git, so git-describe cannot run and 2.1.3 reports -# "unknown" -- in --version and in the clpeak_version field of every JSON/XML/ -# CSV export. Our source is exactly the tag Version names, so state it (Patch0). +# GitHub archives carry no .git, so git-describe cannot run and the build would +# report "unknown" -- in --version and in the clpeak_version field of every +# JSON/XML/CSV export. Our source is exactly the tag Version names, so state it. %cmake \ -DCLPEAK_VERSION_OVERRIDE=%{version} \ -DCLPEAK_ENABLE_GUI=OFF \ ++++++ clpeak-2.1.3.tar.gz -> clpeak-2.1.4.tar.gz ++++++ ++++ 8920 lines of diff (skipped)
