qazal
f9cfb64cd9
test asm_gemm in CI ( #14551 )
...
* test asm_gemm in CI
* default float16
* use a smaller shape for multi
* smaller size
* smaller for CI
* smaller for ci
* need half
2026-02-05 13:32:22 +09:00
Christopher Milan
232848d086
PYTHONREMU: VOP3P integer operations with constants don't cast to fp16 ( #14546 )
...
* PYTHONREMU: VOP3P integer operations with constants don't cast to fp16
* put that back
* cleaner
* do that once
2026-02-04 20:10:59 -05:00
wozeparrot
2966619834
feat: llama uses enable_gqa during training ( #14545 )
2026-02-04 16:22:31 -08:00
Christopher Milan
5338ce6b74
test S_PACK in extra/assembly/amd/test/hw ( #14537 )
...
* S_PACK_LL_B32_B16 in test/hw
* add rest of S_PACK instructions
2026-02-04 14:17:16 -05:00
chenyu
9052db678f
remove allow_shape_mismatch in Tensor.replace ( #14536 )
...
move all logic to torch_backend and not hacking Tensor method
2026-02-04 12:38:18 -05:00
nimlgen
62786d488a
am: mi3xx perf ( #14529 )
2026-02-04 19:32:43 +03:00
chenyu
d57d24c7d4
Buffer.as_buffer -> Buffer.as_memoryview [pr] ( #14535 )
...
it casts to memoryview. also inline the as_typed_buffer checks to Tensor._data
2026-02-04 11:31:11 -05:00
chenyu
67f91e897b
UOp.is_contiguous -> UOp.has_buffer_identity [pr] ( #14530 )
...
one more confusing buffer related method, but it's definitely not is_contiguous
2026-02-04 09:21:26 -05:00
Christopher Milan
ecbce5269e
PYTHONREMU properly supports S_PACK_LL_B32_B16 ( #14527 )
...
* PYTHONREMU properly supports S_PACK_LL_B32_B16
* default
2026-02-03 23:45:33 -05:00
wozeparrot
720c9597a9
feat: llama uses is_causal on sdpa during training ( #14528 )
2026-02-03 20:24:30 -08:00
qazal
d1bfbe9ce3
isolate slow llama gemm ( #14525 )
2026-02-04 12:20:10 +09:00
George Hotz
d59e6e7a37
move more tests to test/null, split some existing ones ( #14512 )
...
* move more tests to test/null, split some existing ones
* null work
* null work
* move more
* fixes
* move PIL
* PIL in CLIP
* don't move that
2026-02-03 20:20:20 +08:00
qazal
a98c53769a
ASM_GEMM=1 runs the UOp gemm on non cdna ( #14516 )
...
* ASM_GEMM=1 runs the UOp gemm on non cdna
tests run on mac in 3 seconds
* min diff
2026-02-03 20:42:02 +09:00
qazal
5c1d21349e
viz: profiler command line tool ( #14515 )
2026-02-03 19:51:25 +09:00
George Hotz
dd2de4f838
rename all DEFINE_GLOBAL to PARAM ( #14511 )
2026-02-03 15:09:38 +08:00
wozeparrot
bbcd3d67a3
fa: faster ( #14453 )
2026-02-02 21:34:17 -08:00
chenyu
66d2b02f11
delete files that depends on extra.optimization.helpers ( #14499 )
2026-02-02 13:33:33 -05:00
George Hotz
6e958dbfd4
assembly/amd: add RDNA4 support to emulator ( #14341 )
...
* start new rdna4
* work
* plus works
* more pass
* rdna4
* assembly/amd: fix RDNA4 emulator for float16 and VOP3 clamp
* stale
* rev
* rr
* rdna4 emu tests
* cleanup
* cleanup
* simp
* works
* better factorizaion
* hacks
* fix mockgpu
* guard both
* cleaner
* gate
* bug fix and a few tests
* all test_tiny
2026-02-02 21:35:59 +08:00
qazal
965940dd00
sqtt: update examples after event field change ( #14493 )
...
* regen sqtt examples
* cdna
* rdna4
* packet counts for rdna3
* sqttmap work
2026-02-02 21:39:48 +09:00
George Hotz
965149a46d
assembly/amd: add ds perm instructions ( #14486 )
...
* assembly/amd: add ds perm instructions
* NO SKIP
* fix preexisting RDNA3 issues
* pcode
* assert
* asserts
* unify
* simp
* good fix
2026-02-02 16:02:00 +08:00
Robbe Derks
d75a1b0d5a
usbgpu: use BOT interface for patch.py ( #13644 )
...
* BOT usage
* cleanup
* fix lint
* fix ruff
* fix -7?
2026-02-02 11:54:46 +08:00
qazal
616e9c1483
CDNA assembly gemm in tensor.py with flag ( #14310 )
...
* work
* work
* the assembly
* remove the old one
* remove ws bufs, assert splitk
* notes cleanup
* work
* gemm args
* gemm in mixins would be nice
* add gemm gradient
* print counters
* the realize is for DEBUG=2 aesthetics
* dedup
* rewrite to python dsl, no list copies
* leave that
* add B, M, N, K to gemm name
* it's M0 not NULL
* fp16 support
* test cleanup + more gemms
* work from viz
* more work
* gemm batch_size
* xccg path work
* tiny comments on the label naming
* s_waitcnt
2026-01-31 22:34:14 +09:00
qazal
d69bc5aa1a
make DEV=NULL EMULATE=AMD amd_asm_matmul run ( #14460 )
2026-01-31 20:45:24 +09:00
George Hotz
b705c9143c
assembly/amd: test more instructions ( #14365 )
...
* assembly/amd: test more instructions
* more
* passing
* revert
* no const fold
* remove junk
* cleaner
2026-01-31 12:40:22 +08:00
Christopher Milan
e575dd8275
prevent UB in long decomp and more emulated tests ( #14447 )
2026-01-30 19:38:41 -05:00
nimlgen
1998e0bb28
nv: add prof props to dev ( #14437 )
2026-01-30 12:51:43 +03:00
wozeparrot
c2fb8b208f
fa: 32 block size ( #14416 )
2026-01-29 13:59:13 -08:00
nimlgen
dc977a03b0
nv_pma: bw decoder ( #14424 )
...
* nv_pma: bw decoder
* decoder fix
* better
2026-01-30 00:12:39 +03:00
nimlgen
2d5c24879f
nv: pma for 5090 ( #14420 )
...
* nv: pma for 5090
* hm
* 4090
2026-01-29 20:06:01 +03:00
nimlgen
acb2fc36ba
nv_pma: add decoder ( #14404 )
...
* nv_pma: add decoder
* cl
2026-01-28 20:44:02 +03:00
nimlgen
544928766d
hcq_smi: kill mac pids ( #14398 )
2026-01-28 15:00:28 +03:00
George Hotz
202b74b369
assembly/amd: continue refactors ( #14386 )
...
* simpler
* merge
* flat
* no ctx
* use the correct apis
* dup code
* write clean code
* remove bad helpers
* bits junk remove
* junk remove
* smem test
* fix tests
* correct fix + tests
* Fmt matters it seems
* wmma refactor
* a lil more
* kimi cleanups
* line
2026-01-28 17:33:03 +08:00
qazal
0294014108
fix bufferize cost function for multi, improve VIZ=-1 cli ( #14394 )
...
* improve cli
* remove_bufferize change
2026-01-28 15:53:18 +09:00
George Hotz
88bc5ee212
assembly/amd: rename to better names ( #14384 )
...
* assembly/amd: rename to better names
* might help fuzzing segfault
* emu2 -> emu
2026-01-28 10:00:54 +08:00
chenyu
db010a31be
IGNORE_OOB -> CHECK_OOB [pr] ( #14374 )
...
flip the meaning
2026-01-27 12:20:59 -05:00
George Hotz
e5df7e640b
fix branches in amd_asm_matmul ( #14369 )
2026-01-27 20:48:42 +08:00
George Hotz
bfc88bcfb8
assembly/amd: emu refactors + enable PYTHON_REMU by default ( #14361 )
...
* assembly/amd: start refactors
* cleanups
* those are global
* methods on ctx
* const cleanup
* range helper
* types and imports
* cleanups
* cleanups
* remove stale name
* fix emu2 types
* more typing
* more mypy
* cleanups
* fxns
* scc cleanup
* cleanups
* cleanups
* simpler parse_pcode
* laneid
* no defaults for pcode
* pcode is not optional
* cleanups
* functions cleanup
* splat
* expr_parser functions
* single tok
* invert global loops
* try_eat
* minor
* run parser on all
* no silent 0
* tests
2026-01-27 17:42:24 +08:00
qazal
f866b2a513
mfma loop in asm dsl ( #14349 )
...
* mfma loop in asm dsl
* work
2026-01-27 11:11:37 +09:00
George Hotz
204f51e739
assembly/amd: bug fixes for PYTHON_REMU ( #14347 )
...
* default PYTHON_REMU to 1
* mockgpu
* less size
* normal compile path
* uniqie
* more
* fix clamp
* Change PYTHON_REMU default to 0 in _try_dlopen_remu
2026-01-27 00:48:22 +08:00
George Hotz
3b43d26f10
assembly/amd: emu speed ( #14344 )
...
* assembly/amd: emu speed
* fix spec
* go
* don't do this
* simpler
* no stupid consts
* hack
* simpler
* no index
* no where
* faster linearizer
* fix spec
* no index dtype
2026-01-26 22:21:34 +08:00
George Hotz
774a454bb5
assembly/amd: fix scratch SVE ( #14340 )
...
* assembly/amd: default python REMU
* mem_used
* no lane
* sve
* remove that
* needs s_code_end in tests
2026-01-26 21:03:51 +08:00
qazal
2d91fe6310
use amdgpu dsl in mmapeak ( #14342 )
...
* use amdgpu dsl in mmapeak
* don't rely on llvm for vgpr counting
* llvm roundtrip assert
* rm it, add ci
* vgpr_count
* move emulated test to amd, it needs comgr
* env
* arch
* inst._fields -> inst.operands
* vgpr offset
2026-01-26 22:03:43 +09:00
George Hotz
be23776ba7
assembly/amd: replace pcode with ucode ( #14002 )
...
* a bunch of todos for my boy claude
* uops have types
* lil cleanups
* simpler ucode
* isNAN
* calls
* move more
* cleanup pcode_parse
* cvt functions
* fix parser bugs
* no void
* minmax
* more pcode parse
* pretty print
* transform
* comments
* move to transform
* assign/declare
* simpler norm
* single PM
* just Uops
* simpler
* more typed
* all rewrite
* less verbose
* work
* spec
* transform
* work
* simpler spec
* less spec
* bitcast
* simpler
* simp ucode
* work
* more in pcode_transform
* remove junk
* more functions
* bug
* no void assign
* load/store
* wave
* fixes
* move denorm
* move more functions
* tests
* cat is shape None
* uop syntax
* move a few more
* program_spec
* cat stuff
* assign fix clear
* unused
* nans
* fp bits
* works with simplify
* remove junk
* special
* meh
* more
* more
* update test pcode parse
* improve parser
* parse some for loops
* merge master
* dead files
* tests pass
* emu2
* better emu2
* test_plus works
* uselessly write more instructions
* use pcode
* something
* something
* bench_emu
* progress
* ds works
* work
* work
* more passing
* run compare
* bench_emu
* more pcode
* a few more
* bugfixes
* bugfix
* test fixes
* tests pass without USE_HW
* all hw tests pass
* add more hw tests
* new hw tests
* bit
* less handcode
* parse more
* consolidate pcode
* fixes
* rsrc
* lane pcode
* cleanups
* simpler
* emu bugs
* one cmp test fails
* fix decode and upd name
* fix name and test harness
* _ftz_f32
* fix denorm
* fix VOPD and use load
* fix carry bug
* no load where / just invalid
* clean
* simpler
* merge sops
* refactoring
* simplifications
* bugfixes
* new tests
* f16 sin fix
* assertion and hw tests
* cvt functions
* one more failure
* bugfixes
* bugfix + regression
* more tests
* fmac
* no manual unrolling
* ordering
* LLVM backend is a lot faster
* compile inst
* more bugs
* f16
* bugfix
* fix regression
* one clang call
* 1M inst
* scratch works
* do scratch correctly
* cleanup
* regression
* cmp
* fmamk fixes
* merge
* fix vcmpx
* unify memory
* remove unused code
* ignore oob for test
* cleanups
* fix mbs
* unify cmp
* test
* minor cleanups
* bump timeout
* fix tests
* revert the CMPLE stuff
* remove opt
* less diff
* simpler
* revert
* support multiple backends
* memset is a lot faster
* split out in bench emu
* improve timing
* timing
* cache that
* cache that
* simpler and faster
* tokenize
* binop table
* simpler
* move to parser
* tok for lambda
* refactor
* expr_parser
* delete emu2_pcode
* import cleanup
* lil
* if parse
* work
* simpler
* no v
* trig preop is faster
* durations for tests
* fix cmp bug
* sdst
* remove scartch_size hack
* null behavior
* _MXCSRContext
* bugfixes
* DEBUG >= 3
* test smem crashes my gpu
* debug
* test
* test smem
* profiler
* full inst
* bugfix
* rtag(1)
* pc is 64-bit and word
* pc is real code now
* dynamic
* more dynamic
* fix oob access
* fix crash, more dyn
* all dyn
* really all dyn
* correct null mask
* lit + format
* 21s on the tests
* 13s on the tests
* canonical name
* simm16
* more dyn
* 14s
* proper saddr dedup
* dyn
* debug 5
* better 5
* revert dynamic stuff
* that can be dyn
* negative offsets
* dyn wmma
* f16 wmma support / ops / dtype / dtype_alu
* symbolic changes not needed
* ConstFloat
* more uop.const
* __eq__
* uop tests
* fix f16
* bf16 tensor cores
* whitespace
* remove cast roundtrip
* Revert "remove cast roundtrip"
This reverts commit c5bb0381c3
2026-01-26 18:04:29 +08:00
qazal
92bfe92138
assembly/amd: fix cdna mfma xml ( #14329 )
...
* handwritten failing test
* new amdxml
* more mfma from fixes
* ci
* move arch of test integration
* alt
* amdxml human cleanup
* _TestIntegration rename to IntegrationTestBase
* it's the same problem as _LIT
* better comment
* better variable name
2026-01-26 17:51:26 +09:00
chenyu
e3601788fa
update torch backend function ( #14333 )
...
those have tensor.py implementation
2026-01-25 16:39:34 -05:00
nimlgen
9865f51e39
cupti: ref collector ( #14330 )
...
* cupti: ref collector
* ll
2026-01-25 20:35:21 +03:00
George Hotz
49db266b96
ReprEnum for repr roundtrips ( #14327 )
...
* ReprEnum for repr roundtrips
* dsl
* bugfixes
* vdsty fixes
* cleaner
* fix
* fix cdna fields
* tests all pass
2026-01-25 18:58:31 +08:00
qazal
bf2d9d138f
viz: simplify amdgpu cfg ( #14326 )
...
* viz: replace llvm disasm with our disasm
* it starts with more code
* then it becomes less
* simpler, cdna disassembles with decimal simm16
* s_branch is upper case, add test
* simm16s and others
2026-01-25 15:21:45 +09:00
qazal
647e527a7e
viz: replace llvm disasm with our disasm ( #14325 )
2026-01-25 13:56:56 +09:00
chenyu
7e41da1ae8
fix generate_dataset.sh ( #14324 )
...
added `set -e` so wrong pathes would fail the script, then fixed the path
2026-01-24 16:47:10 -05:00
