| Index | index by Group | index by Distribution | index by Vendor | index by creation date | index by Name | Mirrors | Help | Search |
| Name: clpeak | Distribution: openSUSE Tumbleweed |
| Version: 2.1.2 | Vendor: openSUSE |
| Release: 1.1 | Build date: Thu Aug 20 20:40:38 2026 |
| Group: Unspecified | Build host: reproducible |
| Size: 1105491 | Source RPM: clpeak-2.1.2-1.1.src.rpm |
| Packager: https://bugs.opensuse.org | |
| Url: https://github.com/krrishnarraj/clpeak | |
| Summary: Find peak OpenCL capacities like bandwidth & compute | |
A tool which profiles OpenCL devices to find their peak capacities like bandwidth & compute.
Apache-2.0
* Thu Aug 20 2026 Martin Pluskal <mpluskal@suse.com>
- Update to 2.1.2:
* Keep machine-readable output locale-independent so XML
dumps, JSON/CSV exports and parsed values always use '.'
as the decimal separator, including under comma-decimal
locales (gh#krrishnarraj/clpeak#199).
* Wed Aug 19 2026 Martin Pluskal <mpluskal@suse.com>
- Update to 2.1.1:
* New CLI tests: bilinear texture sample rate
(--texture-sample), store-to-load forwarding latency
(--store-forward), and SMT scaling of fp32 throughput
(--smt-scaling).
* Rebuild compute MAD chains as a single accumulator across
all backends; the old two-register ping-pong understated
fp32 peak on some GPUs. Numbers on the same machine can be
higher than 2.0.19.
* Fix CPU cache-bandwidth measurement (read/write/copy) and
stop the AVX-512 FP16 chain collapsing under fast-math.
* --describe prints what each test and reading measures.
* Overlapping TestScopes no longer silently drop result rows
from the text output.
* Exclude per-ISA CPU translation units from LTO so distro
- flto builds keep their -m/-march flags at link time.
* 2.1.1: drop the AMX-TF32 matrix test (Intel removed the ISA
before any silicon shipped) and probe AMX tile ops at
configure time instead of trusting compiler flags
(gh#krrishnarraj/clpeak#193, gh#krrishnarraj/clpeak#195).
* Upstream also adds a Flutter desktop/mobile GUI; it is not
packaged here (no Flutter SDK in Factory).
- Enable the Vulkan backend (pkgconfig(vulkan), vulkan-headers,
shaderc) and drop the unused pkgconfig(gl) BuildRequires.
- Set the tarball version fallback so clpeak --version reports
2.1.1 rather than the stale 2.0.16
(gh#krrishnarraj/clpeak#198).
* Mon Aug 17 2026 Martin Pluskal <mpluskal@suse.com>
- Update to 2.0.19:
* OpenCL: replace the mad() builtin with a contracted (y*x)+y
expression in the fp32/fp64/fp16 compute kernels. The spec only
permits, but does not require, mad() to fuse into a hardware
FMA, and Apple's CL-on-Metal compiler never lowered it, halving
the reported fp32 throughput there; other platforms are
unaffected because they already fused it.
* CUDA/ROCm: mark the global-bandwidth kernel in/out pointers
__restrict__, so the compiler may batch the loads ahead of the
trailing store and NVIDIA can take the read-only ld.global.nc
path.
* CPU backend: fix the GCC build failure on the AArch64 crypto ISA
extensions (gh#krrishnarraj/clpeak#189). GCC exposes the
AES/SHA-256 NEON intrinsics only under the umbrella +crypto
target attribute, so a translation unit compiled with just +aes
or +sha2 failed to inline them; GCC now gets +crypto
(+crypto+sha3 for the SHA-512 unit) while clang keeps the
granular flags.
- Correct License to Apache-2.0: upstream relicensed away from the
Unlicense in December 2023 and the bundled LICENSE has shipped the
Apache License 2.0 ever since, but the spec still declared
LicenseRef-SUSE-Public-Domain.
* Mon Jul 13 2026 Martin Pluskal <mpluskal@suse.com>
- Update to 2.0.18:
* Major expansion of the native CPU backend with new benchmark
probes for recent x86 and ARM ISA extensions (AVX-VNNI /
AVX10.2 / AMX, SVE/SVE2, AES/SHA crypto, string and UTF-8
validation kernels)
* Wed Jul 08 2026 Martin Pluskal <mpluskal@suse.com>
- Update to 2.0.17:
* OpenCL: clamp the work-group size to each kernel's device
maximum, fixing failures on devices with a small max
work-group size.
* Fix benchmark-correctness bugs where dead-code elimination
folded away the matrix accumulator chains and inflated the
reported throughput by up to 4x (CUDA WMMA/MMA for
FP16/BF16/TF32/INT8/FP64 and Metal simdgroup_matrix).
* CUDA: rework the tensor-core path to native mma.sync tiles and
add a full-rate FP16 kernel.
* Enable joint_matrix benchmarking on Intel Arc A770 (oneAPI).
* Fri Jun 26 2026 Martin Pluskal <mpluskal@suse.com>
- Update to 2.0.16:
* The downstream fix to install the binary into the standard
bindir is now upstream (PR #178); install the LICENSE file via
CMake again.
* Raise the build to C++14 to fix aggregate brace-initialization
on strict clang.
* New GPU support: CUDA Blackwell Ultra (B300/GB300), GB10 and
Jetson Thor.
- Drop clpeak-install-bindir.patch, merged upstream.
* Wed Jun 24 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.0.14:
* Gate AMX intrinsics on 64-bit targets to fix i686 builds
* CPU: expose every supported ISA variant as its own compute
test
* CPU: volatile-seed fp32/fp64 affine coefficients to fix the
SSE2 peak result under -ffast-math
* CPU: use matrix instruction names for ARM matrix engine
labels
* ROCm: fix bf16 and rocWMMA AOT kernel builds on ROCm 6.4,
tighten AOT arch groups, and fix dlopen shim compile errors
* Optionally build driver-only GPU binaries (AOT CUDA/ROCm,
optional GEMM libraries)
- Add clpeak-install-bindir.patch: the upstream packaging rework
started installing the binary into the prefix root (/usr/clpeak);
install it into bindir via GNUInstallDirs instead (sent upstream)
- Drop _service/_servicedata and build straight from the upstream
release tarball
* Sat Jun 20 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.0.13:
* Vulkan: reworked coopmat shaders (independent/explicit accumulator
chains for ILP, specialization constants for the tile shape)
* Treat absent-driver / no-device init errors as success across the
Vulkan/CUDA/ROCm backends (real errors are still surfaced)
* ROCm: int8_dp4a and hipBLASLt FP4 probe stability on gfx942
* OpenCL: create the backend scope before the platform query so the
header is always printed
* CPU backend: print L2 numbers for all cores, mixed-precision kernel
refinements, gate NEON to AArch64, drop the redundant memcpy
bandwidth test
* Removed the Thread dispatch latency (CPU) and Atomic throughput tests
* Thu Jun 04 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 2.0.10:
* Add ROCm / HIP backend
Full AMD GPU backend using the HIP runtime. Includes the.
complete benchmark suite as it becomes available on each GPU.
family:
+ Compute throughput — FP64/FP32/FP16/BF16/INT32 vector kernels
with vector-width ILP variants.
+ INT8 DP4a — sudot4 instruction on CDNA3 / RDNA3-4.
+ Raw MFMA matrix-core peak — FP16, BF16, INT8, FP8 (CDNA
architectures).
+ 2:4 structured-sparse MFMA (smfmac) — FP16, BF16, FP8.
+ FP4 matrix-core (MXFP4) — Instinct MI300+.
+ Native WMMA peak — RDNA3/4, covering FP16, BF16,
FP8 E4M3/E5M2.
+ rocBLAS — FP16, BF16, FP8 GEMM.
+ hipBLASLt — FP8 GEMM.
+ Memory bandwidth — global, local, host-device transfer.
* Add oneAPI / SYCL backend
Intel GPU backend compiled with icpx -fsycl. Requires the
Intel oneAPI Base Toolkit.
+ Compute throughput — FP64/FP32/FP16/BF16/INT32/INT8 vector
kernels with vector-width sweeps.
+ joint_matrix peak — device-queried shapes, VNNI B layout,
TF32; per-device capability diagnostics.
+ oneMKL GEMM — FP16, BF16, FP32 (where supported).
+ Memory bandwidth.
+ Falls back to CPU/accelerator device when no GPU is visible.
* New Benchmarks (CUDA)
+ NVFP4 matrix-core — mma FP4 benchmarks for Ada/Hopper.
+ FP4 2:4 sparse mma.sp — Blackwell (SM 12.0+), with corrected
.scale_vec::4X selectors.
* CLI Changes
+ --verbose - Print backend debug logs: kernel build logs,
API errors, and library exceptions that are suppressed by
default.
* Device selection: now accepts comma-separated index lists.
* Removed flags
+ --cl-platform-name <str>
OpenCL platform name matching removed; use --cl-platform
<index>.
+ --cl-device-name <str>
OpenCL device name matching removed; use --cl-device <index>
+ --int4-packed-compute / --no-int4-packed-compute.
Emulated INT4 benchmark removed.
* Thu May 21 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 2.0.9
* Use pseudo-random data in buffer population to defeat memory
compression.
* Apply pseudo-random data initialisation to every bandwidth test
- Update to version 2.0.8
* New iOS app, refactoring of all GPU backends.
- Update to version 2.0.7
* Major Refactor: Architecture Restructuring, Logger Redesign &
Android Improvements.
* Wed May 13 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 2.0.6:
* CUDA integer4/int8 sparse tests, cleanup.
* Runtime gate for opencl int8 dp test.
* Mon May 11 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 2.0.5:
* [vulkan] reduce work size when CU is not known.
* Zero Vulkan atomic buffers before dispatch.
* Calibrate with single-iteration probes.
* [vulkan] remove local bandwidth v8.
* Increase max-time default to 500 ms.
* Initial commit for time based runs.
* Atomic throughput: bump ATOMIC_REPS 64 -> 256 (M1 was
under-measuring).
* Atomic throughput: cut ATOMIC_REPS 512 -> 64 to avoid GPU
watchdog.
* Vulkan: reduce fp64 dispatch size to avoid GPU watchdog on
RDNA4 + RADV
* Kernel latency: report submit failures instead of bogus
roundtrip times.
* Fri May 08 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 2.0.3:
* Vulkan: fix inf/bogus bandwidth on Adreno (Turnip) + drivers
with broken timers.
* Atomic throughput: cut per-test work and unify metric names
across backends.
* Vulkan: split coopmat dispatch into FP and INT phases.
* Vulkan: expand compute test variants and add new tests from
other backends.
* Thu May 07 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 2.0.2:
* Remove debug signing config from build file.
* Change ZERO_COPY_MULTIPLIER from 5.0 to 3.0.
* [cmake] version - don't use long version on release.
* Thu May 07 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 2.0.0:
New Backends
* Vulkan Compute, bandwidth, atomics, cooperative matrix (tensor
core equivalent).
* CUDA Compute, bandwidth, WMMA/BMMA tensor cores, cuBLASLt
GEMM.
* OpenCL Unchanged baseline; now explicitly selectable via
- -opencl
New Benchmarks
* Floating-point Compute
+ BFloat16 (BF16) throughput.
+ CUDA WMMA tensor cores: FP16, BF16, FP8-E4M3, FP8-E5M2, TF32, FP64
+ CUDA BMMA (binary matrix).
+ Vulkan cooperative matrix: FP16, BF16, INT8, FP8 variants.
+ Metal simdgroup matrix: FP16, BF16.
* Integer Compute
+ INT8 dot-product (DP4a).
+ Packed INT4 (emulated).
+ CUDA WMMA INT8 / INT4.
* GEMM — vendor-SDK peak
+ CUDA: cuBLASLt (FP32, TF32, FP16, BF16, FP8-E4M3, FP8-E5M2,
INT8, INT4).
+ Metal: MPS / MPSGraph (FP32, FP16, BF16).
Memory & Latency
* Local (shared) memory bandwidth.
* Image / texture memory bandwidth.
* Host↔Device transfer bandwidth (Vulkan, CUDA).
* Atomic throughput.
CLI Changes
* Several flags from v1.x have been renamed or removed.
Output Format Changes
* XML — breaking schema change (format_version="2")
v1.x XML files are not accepted as --compare baselines.
Bug Fixes & Improvements
* Unified measurement methodology: batch timing replaces
per-iteration event timing across all backends.
* Dispatch sizes auto-scaled to GPU compute unit / SM count.
* RAII scope guard replaces manual resultScopeBegin() /
resultScopeEnd() calls.
* Baseline comparison via --compare FILE accepts any v2 JSON,
CSV, or XML file.
* Sat Mar 14 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 1.1.7:
* Update patch version to 7. No functional changes.
* Thu Mar 12 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 1.1.6:
* transfer_bandwidth: allow subtest filtering with --testName
* [workflow] add macos build
* android: Add compute_char.cpp and compute_short.cpp
* android: build fix for missing aligned_alloc in API 26
* workflows: Add github build workflows
* Thu Mar 12 2026 Martin Hauke <mardnh@gmx.de>
- Update to version 1.1.6
* transfer_bandwidth: allow subtest filtering with --testName.
* Wed Jun 18 2025 Martin Hauke <mardnh@gmx.de>
- Update to version 1.1.5:
* Update common.cmake
* Add option to force number of iteration per kernel
* Tue Dec 24 2024 Martin Hauke <mardnh@gmx.de>
- Update to version 1.1.4:
* Add Tesla T4 results
* Create GeForce_GTX_1660_Ti.log
* Tue Nov 26 2024 Martin Hauke <mardnh@gmx.de>
- Update to version 1.1.3
* Add a few more parameters.
* Add 24-bit integer compute to RX 5500XT results.
* [cmake] use GNUInstallDirs to get canonical paths.
* Add log files for additional devices.
* Fix compute integer unit reported in xml files.
* add profiling for char & short integer types.
* Adding results for Tesla P40.
* Adding results for MTT S80.
* Use stricter host buffer alignment (64B) required by modern
CPUs.
* Wed May 03 2023 Martin Pluskal <mpluskal@suse.com>
- Update to version 1.1.2:
* [cmake] Bump patch version
* [snap] fix yaml for latest version of snapcraft
* [snap] move to core20
/usr/bin/clpeak /usr/share/doc/packages/clpeak /usr/share/doc/packages/clpeak/README.md /usr/share/doc/packages/clpeak/results /usr/share/doc/packages/clpeak/results/AMD /usr/share/doc/packages/clpeak/results/AMD/Ryzen_Threadripper_PRO_3955WX_16_Cores.xml /usr/share/doc/packages/clpeak/results/Apple /usr/share/doc/packages/clpeak/results/Apple/M1_Pro.xml /usr/share/doc/packages/clpeak/results/Nvidia /usr/share/doc/packages/clpeak/results/Nvidia/GeForce_RTX_5060.xml /usr/share/licenses/clpeak /usr/share/licenses/clpeak/LICENSE
Generated by rpm2html 1.8.1
Fabrice Bellet, Sun Aug 23 00:12:05 2026