Rebase/cherry-pick xsmm-main on top of triton-lang/triton-cpu/main#49

Open

adam-smnk wants to merge 407 commits into

libxsmm:xsmm-mainfrom

adam-smnk:rebase-on-main-25-Jan-09

This pull request is big! We're only showing the most recent 250 commits

Commits on Nov 21, 2024

Commits on Nov 22, 2024

[BACKEND] Fix transpose optimization missed during refactor (#5226)
ThomasRaoux
authored
[AMD] Use warp shuffle for fp8 MFMA to dot operand layout conversion (#5139)
ilia-cher
authored
[LAYOUTS] [BE] Simplify Ampere/Hopper paths introduced in #5189 (#5200)
lezcano
authored
[BACKEND] Use LL to simplify redundant elements check and fix related issues (#5225)
Jokeren
authored
Make TMA tests compatible with older CUDA toolchains (#5221)
embg
authored
[CMake] Add C as project language (#5217)
makslevental
authored
[AMD] Fix slow compilation due to inlining print calls (#5153)

binarman
and
antiagainst
authored
[AMD] Re-enable overflow test in test_reduce_layouts (#5233)
antiagainst
authored
[BACKEND] Fix a missed transpose optimization during refactor (#5236)
ThomasRaoux
authored
Revert "Allow Layouts to propogate to local_load" (#5237)
ThomasRaoux
authored
Revert "[AMD] Use warp shuffle for MFMA to Dot operand layout conversion (FP8)" (#5240)
ThomasRaoux
authored
Updated README.md to show the steps for overriding kernel's IR (#5239)
arakhmati-openai
authored
Ensure device context before launching kernel (#3731)
bertmaher
authored

Commits on Nov 23, 2024

Commits on Nov 25, 2024

Commits on Nov 27, 2024

Commits on Nov 28, 2024

Commits on Nov 29, 2024

Commits on Nov 30, 2024

Commits on Dec 1, 2024

Commits on Dec 2, 2024

Commits on Dec 3, 2024

Commits on Dec 4, 2024

Commits on Dec 5, 2024

Commits on Dec 6, 2024

[Pipeliner] Fix a dangling insertion point issue. (#5352)
htyu
authored
Short preamble for the README, explaining why this clone exists

bertmaher
authored and
ienkovich
committed
OSS Automated Fix: Addition of Code of Conduct (#1)

facebook-github-bot
authored and
ienkovich
committed
[BACKEND][CPU] Initial plumbing for cpu backend (#2)

minjang
authored and
ienkovich
committed
[BACKEND][CPU] Create TritonCPU and conversion dialects (#3)

minjang
authored and
ienkovich
committed
Update README.md

minjang
authored and
ienkovich
committed
Convert tt.func and tt.return (#4)

minjang
authored and
ienkovich
committed
[BACKEND][CPU] Convert tt.get_program_id and tt.print (Hello World) (#1)

minjang
authored and
ienkovich
committed
Quick patches to make it work after rebasing (#3)

minjang
authored and
ienkovich
committed
Support basic lowering through vector dialect in CPU backend.
ienkovich
committed
Revert unreviewed changes. (#5)

authored and
ienkovich
committed
Add a workaround for LLVM bug in codegen for bf16 vector cast. (#4)
ienkovich
committed
Prototype of the Triton CPU backend with basic compilation and execution flows (#6)
ienkovich
committed
Add support for tl.cat operation. (#9)
ienkovich
committed
[BACKEND][CPU] Make it buildable and runnable in a different environment (#8)

minjang
authored and
ienkovich
committed
Add support for simple reductions. (#10)
ienkovich
committed
Support tl.histogram for CPU. (#12)

ienkovich
and
minjang
committed
Fix merge and compile errors (#13)

minjang
authored and
ienkovich
committed
[CPU] Support flexible active driver + update vector-add tutorial (#11)

minjang
authored and
ienkovich
committed
Added a simple workflow to run on self-hosted intel runner (#16)

gshimansky
authored and
ienkovich
committed
Fixed build and test workflow for intel self-hosted runner (#17)

gshimansky
authored and
ienkovich
committed
[CPU] Add an OpenMP-based CPU launcher (#15)

minjang
authored and
ienkovich
committed
Support generic reduction and scan cases. (#14)
ienkovich
committed
[CPU] Dump human-readable asm code in TRITON_CACHE_DIR (#19)

minjang
authored and
ienkovich
committed
Added g++ installation after switching to ubuntu-22.04 (#21)

gshimansky
authored and
ienkovich
committed
Support atomic ops for CPU. (#20)
ienkovich
committed
[TUTORIAL] Add unmasked matrix multiply example to triton-cpu (#23)

Kuigesi
authored and
ienkovich
committed
Update matrix-multiplication-cpu tutorial, use preallocated output buffer for CPU. (#24)

Kuigesi
authored and
ienkovich
committed
Fixes for x86 CI workflow (#26)

ienkovich
and
karupayun
committed
Use static compilation for kernels. (#29)
ienkovich
committed
Move byte manipulation ops from elwise ops conversion. (#28)
ienkovich
committed
[TUTORIAL] Add the non-persistent softmax and make it for CPU (#22)

minjang
authored and
ienkovich
committed
Enable few more core tests for CPU. (#31)
ienkovich
committed
Support tt.split for CPU. (#30)
ienkovich
committed
[BACKEND][CPU] Make the CPU backend buildable and runnable in Mac M1. (#18)

Kuigesi
authored and
ienkovich
committed
[CPU] Add conversion for unsupported BF16 ops via target-specific stage (#27)
ienkovich
committed
Enabled simple build&test workflow, disabled old Integration Tests workflow because it fails to run (#33)

gshimansky
authored and
ienkovich
committed
[BACKEND][CPU] Specify CPU target to native for GNU/Linux Arm (#34)

Kuigesi
authored and
ienkovich
committed
Add conversions for mixed precision matmuls. (#32)
ienkovich
committed
[Op support] Support 'get_num_programs' (#39)

Devjiu
authored and
ienkovich
committed
Add fast-math option: allow fp reduction reassociation

Kuigesi
authored and
ienkovich
committed
Change the lowering option for vector.multi_reduction from InnerParallel to InnerReduction.

Kuigesi
authored and
ienkovich
committed
Fix: TrapUnreachable is not controled by fast-math, we set it unconditionally

Kuigesi
authored and
ienkovich
committed
[so] Compile asm to .so as part of staged lowering (#53)

int3
authored and
ienkovich
committed
Add libdevice for CPU. (#52)
ienkovich
committed
[Op support] Dot3D support (#43)

Devjiu
authored and
ienkovich
committed
Support FP8 conversions for CPU. (#40)

ienkovich
and
minjang
committed
[CPU] Support device_print for scalar types first (#54)

minjang
authored and
ienkovich
committed
[TUTORIAL] Add matrix vector multiplication tutorial (#46)

Kuigesi
authored and
ienkovich
committed
Fix FuncOp lowering. (#61)
ienkovich
committed
[CPU] Easy: remove the old initial boilerplate code (#59)

minjang
authored and
ienkovich
committed
[Scf If types] Support conversion of types for scf::if (#45)

Devjiu
authored and
ienkovich
committed
[WA for fp16 torch.matmul] Replace torch.matmul with np.matmul (#44)

Devjiu
authored and
ienkovich
committed
[cpu] Have MulhiUI lowering support scalars (#64)

int3
authored and
ienkovich
committed
[cpu] Fix formatting (#65)

int3
authored and
ienkovich
committed
[cpu] Support tl.load(..., padding="nan") (#69)

int3
authored and
ienkovich
committed
[cpu] Use helpers from OptCommon.h to simplify code (#67)

int3
authored and
ienkovich
committed
[cpu] Follow up to #69 (#70)

int3
authored and
ienkovich
committed
[cpu] Add runtime library for CPU kernels (#73)

int3
authored and
ienkovich
committed
[FP8 tests] Enable several fp8 tests (#49)

Devjiu
authored and
ienkovich
committed
[cpu] Make runtime library build on Linux too (#75)

int3
authored and
ienkovich
committed
[cpu] Get more of test_random.py working (#77)

int3
authored and
ienkovich
committed
[FIX Pytest] Resolve 'importlib' issue (#78)

Devjiu
authored and
ienkovich
committed
Fix importlib issues (#80)

int3
authored and
ienkovich
committed
[cpu] Add test_annotations.py to CI (#81)

int3
authored and
ienkovich
committed
Reduce/disable some tests on CPU for faster CI runs. (#83)

ienkovich
and
Devjiu
committed
[cpu] Don't reuse shuffle dummies (#88)

int3
authored and
ienkovich
committed
Utilize vector math functions from libmvec. (#55)

ienkovich
and
minjang
committed
Make tl.debug_barrier() a no-op on CPU (#89)

int3
authored and
ienkovich
committed
ConvertMemoryOps should not use cf dialect (#91)

int3
authored and
ienkovich
committed
Remove registered pipelines in favor of explicit lists in python. (#93)
ienkovich
committed
Don't use cf dialect in ConvertAtomicOps (#94)

int3
authored and
ienkovich
committed
atomic_rmw ops should return original value (#95)

int3
authored and
ienkovich
committed
Compute a scalar pointer for vector load instead of extracting it from a tensor (#92)
ienkovich
committed
Add pass to optimize masked loads and stores. (#96)
ienkovich
committed
Fix incorrect casts in mask optimization. (#101)
ienkovich
committed
Add conversion for scf.while (#103)

int3
authored and
ienkovich
committed
[TUTORIAL] Add bf16 matrix vector multiplication tutorial (#90)

Kuigesi
authored and
ienkovich
committed
Add an option to use sleef instead of libmvec. (#104)
ienkovich
committed
Enable fast math by default. (#108)
ienkovich
committed
Add more libdevice lowerings (#97)

int3
authored and
ienkovich
committed
Enable rsqrt and floor for BF16. (#109)
ienkovich
committed
Remove specific dwarf version from -g option. (#110)
ienkovich
committed
Enable `min_dot_size`

Devjiu
authored and
ienkovich
committed
[Formatting] Apply formating

Devjiu
authored and
ienkovich
committed
Remove is_cpu arg from do_bench. (#113)
ienkovich
committed
Enable few more tutorials for CPU (#114)
ienkovich
committed
Pass device type to do_bench in autotuner. (#115)
ienkovich
committed
Fix indices extraction from block pointer. (#116)
ienkovich
committed
[cpu] Rework device_print with triton_cpu.print and 1D vector printing (#99)

authored and
ienkovich
committed
[Pytests] Add several suits (#106)

Devjiu
authored and
ienkovich
committed
Identify dot product pattern (mul followed by a sum) for bf16, and convert it to dot product intrinsics (#56)

Kuigesi
authored and
ienkovich
committed
Add optional packing for converting bf16 dot product. (#118)

Kuigesi
authored and
ienkovich
committed
Add load/store scalarization through loops. (#119)
ienkovich
committed
Fix typo. (#122)
ienkovich
committed
Add lit tests for load/store scalarization. (#121)
ienkovich
committed
[cpu][easy] Fix compiler error on clang (#120)

minjang
authored and
ienkovich
committed
Offload a part of masks optimization to the canonicalizer.
ienkovich
committed
Implement get_module_map for cpu backend

int3
authored and
ienkovich
committed
Make test_hooks CPU-compatible by using get_device_key (#128)

int3
authored and
ienkovich
committed
Make CPU runtime lib lookup work for Python 3.8 (#129)

int3
authored and
ienkovich
committed
Implement device_assert (#126)

int3
authored and
ienkovich
committed
Implement isnan, isinf, signbit (#127)

int3
authored and
ienkovich
committed
Vendor sleef as a submodule (#130)

int3
authored and
ienkovich
committed
Add test_debug_dump.py to CI (#131)

int3
authored and
ienkovich
committed
Refactor MathToLibmvec pass (#135)

int3
authored and
ienkovich
committed
[CPU] Add unit test for print with isSigned and several fixes (#132)

minjang
authored and
ienkovich
committed
Refactor math tests + select vector lib backend via kernel option (#136)

int3
authored and
ienkovich
committed
Vectorize expm1, sqrt, and floor using sleef (#137)

int3
authored and
ienkovich
committed
Fix infinite optimization loop for mask optimization. (#138)
ienkovich
committed
Implement libdevice.trunc (#140)

int3
authored and
ienkovich
committed
Remove old LLVM bug workaround. (#141)
ienkovich
committed
Add kernel execution time measurement using hooks for do_bench (#139)
ienkovich
committed
Use llvm_unreachable in cpu_runtime.cpp (#145)

minjang
authored and
ienkovich
committed
Fix undefined symbole error in libTritonCPURuntime.so (#146)

minjang
authored and
ienkovich
committed
[Dot3D test] Enable with lower block size (#117)

Devjiu
authored and
ienkovich
committed
Add an option to choose between default reduction lowering and our own. (#98)
ienkovich
committed
Fix regressions due to rebasing to the latest upstream

minjang
authored and
ienkovich
committed
Update build-test.yml for pybind11

minjang
authored and
ienkovich
committed
[FP8 support] Enable Float8 tests failed after rebase (#151)

Devjiu
and
ienkovich
committed
Use 1-D vector reduction op to convert reduce op (#152)

Junyi Mei
authored and
ienkovich
committed
[Keep materialization] Turn on meterialization (#154)

Devjiu
authored and
ienkovich
committed
[Scalarization/Loops generation] Refactor and new pass/interfaces introduced (#123)

Devjiu
authored and
ienkovich
committed
Lower memory ops with vector gather and scatter (#158)

Junyi Mei
authored and
ienkovich
committed
Introduce DotOp lowering to AMX (#157)
ienkovich
committed
Implement more libdevice functions using extern_elementwise (#161)

int3
authored and
ienkovich
committed
Fix compilation when ARCH_REQ_XCOMP_PERM isn't defined (#163)

int3
authored and
ienkovich
committed
[CPU] Drop MLIR prefix in ScalarizeInterface (#164)

minjang
authored and
ienkovich
committed
Pad size 2 vectors to size 4 when lowering extern_elementwise ops (#162)

int3
authored and
ienkovich
committed
Rebase onto upstream triton ff306da26b and fix regressions

minjang
authored and
ienkovich
committed
Simple fixes to build on MacOSx (#165)

digantdesai
authored and
ienkovich
committed
Fix trailing null char in ulpSuffix (#166)

digantdesai
authored and
ienkovich
committed
Rebase onto upstream triton 4a5431159a and fix regressions

minjang
authored and
ienkovich
committed
[Test][Autotuner] Skip use_cuda_graph for non cuda devices (#169)

Devjiu
authored and
ienkovich
committed
Add num_threads option to control threading per kernel invocation. (#170)
ienkovich
committed
[TTC Print Memref] Simplify further multidimensional tensor printing (#160)

Devjiu
authored and
ienkovich
committed
Small fixes for autotuner on CPU (#172)
ienkovich
committed
Small fixes for clang + macosx (#173)

digantdesai
authored and
ienkovich
committed
Support multi-dimensional tensor prints in CPU runtime. (#174)
ienkovich
committed
Fix linux-aarch64 build (#176)

desertfire
authored and
ienkovich
committed
Fix math tests for armv8 (#178)

digantdesai
authored and
ienkovich
committed
Allow using local omp with Apple clang (#181)

digantdesai
authored and
ienkovich
committed
Add pytest.mark.cpu to two more already-passing tests (#183)

int3
authored and
ienkovich
committed
Move libdevice to third_party (#182)

int3
authored and
ienkovich
committed
Introduce triton_cpu.DotOp.
ienkovich
committed
Fixes to use the latest LLVM.
ienkovich
committed
Fix pybind11 build issue for TritonCPU.
ienkovich
committed
Use mlir::amx::TileType.
ienkovich
committed
Fix formatting
ienkovich
committed
Fix test_tl_range.
ienkovich
committed
Fix test_conversions.
ienkovich
committed
Disable test_block_copy with lower bound check.
ienkovich
committed

Commits on Dec 9, 2024

Fix isSigned and add float16 in PrintOp (#191)
minjang
authored

Commits on Dec 10, 2024

Add TritonCPU canonicalizer. (#192)
ienkovich
authored

Commits on Dec 12, 2024

Commits on Dec 17, 2024

Fix extra-store in matmul tutorial. (#198)
ienkovich
authored

Commits on Dec 19, 2024

Remove unnecessary bounds checks. (#199)
ienkovich
authored

Commits on Dec 21, 2024

Enable armv8 CI (#195)
digantdesai
authored

Commits on Dec 23, 2024

Fix isSigned usage for scalar prints. (#201)
ienkovich
authored

Commits on Jan 9, 2025

Rebase/cherry-pick our work on top of triton-lang/triton-cpu/main (#30)

committed
Add --backend option and run experiments one backend*config option at a time

Rolf Morel
authored and
adam-smnk
committed
[Triton-benchmark] Set rep=1000 (vs default of 100)

Rolf Morel
authored and
adam-smnk
committed
Fix debug sizes still being set

Rolf Morel
authored and
adam-smnk
committed
Fix typo: benchmark -> config

Rolf Morel
authored and
adam-smnk
committed
Add raw-to-csv.py converter script (#29)

Rolf Morel
authored and
adam-smnk
committed
Make raw-to-csv.py compatible with BENCHMARK-outputting script

Rolf Morel
authored and
adam-smnk
committed
Fix errant call to do-nothing "preproccess" function on some configs (#43)

Rolf Morel
authored and
adam-smnk
committed
Improve raw-to-csv a la @rengoline and prepare for summary line to be parsed

Rolf Morel
authored and
adam-smnk
committed
Change back max dim size to 3k for data collection

Rolf Morel
authored and
adam-smnk
committed
Error out when attempting run bf8 with torch

Rolf Morel
authored and
adam-smnk
committed
Store commit hash with benchmark run details

Rolf Morel
authored and
adam-smnk
committed
Update xsmm_utils.cpp to avoid using libxsmm for now

alheinecke
authored and
adam-smnk
committed
Fix broken torch.compile

Rolf Morel
authored and
adam-smnk
committed