Dean
3d9ba7fce1
Merge pull request #7171 from MrMha/fix/oracle-cli-input-errors-20261017
...
fix: report oracle CLI input errors before running tools
2026-10-18 14:50:08 +13:00
Dean
f599860471
Merge pull request #6770 from MrMha/fix/oracle-path-compiler-20261016
...
fix: key oracle builds by the compiler selected through PATH
2026-10-18 12:56:57 +13:00
MrMha
f83b6c0ee0
fix: report oracle CLI input errors before running tools
2026-10-17 18:41:57 +04:00
Dean
94e67bd135
Merge pull request #6433 from MrMha/fix/oracle-output-freshness-20261015
...
fix: require fresh exact-size output for every oracle dispatch
2026-10-18 02:15:39 +13:00
Adrià
40255d774b
Merge pull request #6081 from MiniRML/fix/oracle-device-selection-cache
...
fix(tools): cache oracle targets per visible device selection
2026-10-17 00:18:07 +02:00
MrMha
7e383a8943
fix: key oracle builds by the compiler selected through PATH
2026-10-16 15:02:22 +04:00
MrMha
9ab52fefc4
fix: require fresh exact-size output for every oracle dispatch
2026-10-15 18:57:14 +04:00
MiniRML
a76fcc6648
test(tools): support Python 3.9 in oracle cache regression
2026-10-14 16:25:47 -07:00
MiniRML
0f0aecea85
fix(tools): cache oracle targets per visible device selection
2026-10-14 14:32:08 -07:00
MiniRML
99f7f6fe06
fix(tools): reject empty hardware oracle input files
2026-10-14 14:22:51 -07:00
lorkydey
af66601f41
feat(hw-oracle): compare captured results bit for bit
2026-10-14 15:53:11 +02:00
lorkydey
0c62cd4a66
feat(hw-oracle): export results with run settings as JSON
2026-10-14 04:06:05 +02:00
Adrià
39890031e3
Merge pull request #4591 from MiniRML/fix/oracle-input-rows
...
fix(oracle): validate input words before tool lookup
2026-10-13 17:25:30 +02:00
MiniRML
17a3d8a1a0
fix(oracle): honor current target override
2026-10-12 09:27:33 -07:00
MiniRML
521cc95b18
fix(oracle): validate input words before tool lookup
2026-10-12 09:26:05 -07:00
MrMha
cf1195ad20
fix(hw-oracle): narrow cache PR and normalize runtime paths
2026-10-11 19:31:39 +04:00
MrMha
563b4738a5
fix(hw-oracle): invalidate cached builds when inputs change
2026-10-11 18:24:42 +04:00
Adrià
f4b7151097
Merge pull request #2510 from Edd016/fix/hw-oracle-mingw-lld
...
fix(hw-oracle): select the AMDGPU ELF emulation when linking kernels
2026-10-11 11:33:58 +02:00
MarcZX100
70d4903158
fix(hw-oracle): batch wave64 probes into 512-row dispatches
2026-10-10 18:28:46 +00:00
ed and Claude Opus 5.5
a83bfa6813
fix(hw-oracle): select the AMDGPU ELF emulation when linking kernels
...
LLD built for a MinGW host runs the COFF linker when ld.lld gets no -m option,
so it rejects the AMDGPU object with "unknown file type" and the
hardware_oracle test fails on Windows with WinLibs LLVM in PATH. Passing
-m elf64_amdgpu selects the ELF linker on every host.
Co-Authored-By: Claude Opus 5.5 <[email protected]
2026-10-09 17:09:53 -07:00
ale and Claude Fable 5.1
8d30c27589
feat(hw-oracle): make the group segment size an option
...
The template allocated a fixed 4 KiB of LDS, so kernels with immediate
offsets past it could not run. --lds (Python lds) sets the group segment
size in bytes, up to 64 KiB, with 4096 as the default.
Co-Authored-By: Claude Fable 5.1 <[email protected]
2026-10-09 12:47:59 -03:00
Dean Galvin
e0a8297a7b
fix(hw-oracle): skip the wave64 assembly test when clang has no AMDGPU target
2026-10-09 23:05:41 +13:00
Dean
b7218b3cec
Merge pull request #1962 from CouchRiv/fix-hw-oracle-wave64
...
fix(hw-oracle): assemble wave64 bodies in wave64 mode
2026-10-09 22:43:05 +13:00
Alex and Claude Opus 5.5
dc902b2eae
fix(hw-oracle): assemble wave64 bodies in wave64 mode
...
clang assembled every body as wave32, so a --wave64 body that used vcc
or an SGPR pair as a condition failed with "operands are not valid for
this GPU or mode". Pass -mwavefrontsize64 on wave64 runs.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]
2026-10-08 23:03:22 -04:00
SK Arin
466ebde23c
fix(oracle): publish cached executable only after successful build
2026-10-08 17:16:36 +05:30
Muhammad Umair
f2b6b5a1a9
fix(oracle)!: require explicit float modes
...
BREAKING CHANGE: CLI and Python callers must supply all seven float
controls. Add clamp, rounding and FP16 overflow settings so hardware
measurements cannot silently use an unintended mode.
Refs #981
2026-10-07 14:58:28 +05:00
Adrià Franch and Claude
f617b40a45
feat(tools): add a hardware oracle for RDNA instruction semantics
...
Co-Authored-By: Claude <[email protected]
2026-10-06 13:54:28 +02:00
Dean