qazal
78a28227c6
assembly/amd: cdna4 mfma support ( #14206 )
2026-01-21 09:12:05 +09:00
George Hotz
1baefed530
assembly/amd: add hw tests from ucode branch ( #14259 )
...
* assembly/amd: add hw tests from ucode branch
* fix is per lane
2026-01-21 08:53:54 +09:00
Robbe Derks
c7fbd177d4
USBGPU: debug script for comma chestnut ( #14252 )
...
* initial debug script
* improvements
2026-01-20 18:52:25 +03:00
nimlgen
dc82856084
tbgpu: shim binary + remote apl pci dev ( #14124 )
...
* shim binary + remote pci dev
* v2
* rip out apl
* cmds
* rename
* clean
* remove
* rm gitignore
* ui
* install
* linter
* um
* cleaner
* assets
* normal install in ui
* cleaner app
* install script
* support fd mmap
* cleaner
* kill server when disconn
* rename + pcidevs
* sign
* install and reinstall
* no sip install
* will trigger update
* nv
* ugh
* this
* fix
* nv
* use nosip sign
* auto install
* remove
* mypy
* upd
* ditto
* print
* simpler
* ditto
* um
* simpler
* upd
* upd
* cleaner
* autogen
* cleaner
* move
* annotations
* server cleaner
2026-01-20 16:15:18 +03:00
qazal
4548fcc1b8
amd/sqtt: add rdna4 and cdna sqtt examples ( #14251 )
...
* amd/sqtt: add rdna4 and cdna sqtt examples
* work
* comment out rdna and cdna tests
2026-01-20 21:11:48 +09:00
qazal
2dc281b32a
assembly/amd: test helpers for arch to gfx target mapping ( #14250 )
2026-01-20 20:35:09 +09:00
George Hotz
0243f4a0f1
clear wins from ucode branch ( #14243 )
...
* clear wins from ucode branch
* two more
* revert those
2026-01-20 15:11:09 +09:00
wozeparrot
1f89eaf790
tk: fa bert mask fix + some numerical stability improvements ( #14214 )
2026-01-19 19:18:07 -08:00
nimlgen
7cb7abeeb0
amd: fix scratch_wave64_lane_byte_size ( #14223 )
2026-01-19 15:21:39 +03:00
qazal
e27a0002c5
viz: only keep the sqtt bytes for pkts ( #14203 )
...
* viz: only keep the sqtt bytes for pkts
* better option name
* work
* renames
2026-01-18 17:04:26 +09:00
qazal
d8f87ae2f2
SQTT packets to assembly mapper ( #14198 )
...
* disasm + compare to llvm
* start inst trace
* base tests pass
* work
* work
* all kernels
* qol
* refactor
* work
* work
* wave_focus
* simple
* work
* add a lot of asserts
* focus on wave0
* correct handling of IMMEDIATE_MASK
* work
* viz work
* use the metadata infra
* better
2026-01-18 16:32:13 +09:00
George Hotz
a51e0a86db
assembly/amd: clean up disasm.py + add CDNA support ( #14200 )
...
* assembly/amd: clean up disasm.py
* cleanups
* add missing encodings
* decode is pretty
* cdna
* assert on failure
* cdna roudtrip
* cdna passing
* test
* lil cleanup
* variant cleanups
* cleanups
2026-01-18 14:48:44 +09:00
George Hotz
79c1559f69
amd asm can still be simpler ( #14199 )
...
* amd asm can still be simpler
* simpler
* V_LANE_ID
* simpler
* simpler
* compact vgpr
2026-01-17 18:40:10 +09:00
George Hotz
50554115ee
fix VALU_SALU / IMMED_MASK and improve amd_asm_matmul ( #14196 )
...
* fix VALU_SALU / IMMED_MASK and improve amd_asm_matmul
* immed
* wave override
* restore ALT
* advance sgprs correctly
* no helpers
* decrease to 192 VGPRs
2026-01-17 11:58:34 +09:00
wozeparrot
a879b54234
tk: fa jit fix ( #14170 )
2026-01-16 16:38:45 -08:00
George Hotz
8a2549d42b
improve amd_asm_matmul + minor VIZ PKTS improvements ( #14186 )
...
* improve amd_asm_matmul + minor VIZ PKTS improvements
* fix waitcnt issue
* cleanups
2026-01-17 06:56:59 +09:00
George Hotz
7d1d9d4568
assembly/amd: remove IMG instruction support and asm.py ( #14163 )
...
* assembly/amd: return IMG instruction supports
* remove asm.py
* op2dsl
2026-01-17 06:21:50 +09:00
nimlgen
e855ec8ee3
tbgpu: refactor dext to support user mappings ( #14177 )
2026-01-16 15:55:57 +03:00
nimlgen
a0dd9d2146
tbgpu: correct com.apple.developer.driverkit.transport.pci entitlements ( #14164 )
...
* tbgpu: correct com.apple.developer.driverkit.transport.pci entitlements
* format
2026-01-15 20:56:39 +03:00
Christopher Milan
0cb024a5bb
remove ctypes.Structure ( #13651 )
2026-01-15 05:06:22 -05:00
George Hotz
255e0573b1
assembly/amd: clean up asm/disasm ( #14158 )
...
* assembly/amd: clean up asm/disasm
* update disasm
* revert dumb stuff
* update decode
* use fmt
2026-01-15 17:45:40 +09:00
qazal
b46da603fe
codegen/custom_kernel: do not attach KernelInfo to user program ( #14160 )
2026-01-15 14:01:48 +09:00
George Hotz
fd60626ea1
assembly/amd: refactor to use op_bits/op_regs ( #14156 )
...
* assembly/amd: refactor to use op_bits/op_regs
* remove that skip
* remove another hack
* remove another hack
* precompute mask
* more reg, less hasattr
2026-01-15 11:20:21 +09:00
George Hotz
e9ce12028e
assembly/amd: amdxml cleanups, remove broken SDWA/DPP, merge in pdf.py ( #14154 )
...
* assembly/amd: amdxml cleanups, remove broken SDWA/DPP
* remove buf junk
* simplify
* simplify
* lil cleanup
* dead fixes
* strip non pcode extraction from pdf
* merge pdf.py into amdxml.py
* only amdxml
2026-01-15 09:23:19 +09:00
wozeparrot
7e5687f6a3
more fa multi fix ( #14152 )
2026-01-14 13:57:11 -08:00
chenyu
986e865830
fix TINY_BACKEND=1 cumsum ( #14138 )
...
* fix TINY_BACKEND=1 cumsum
old hack was wrong, need to apply contiguous on the input
* test time
* test_linalg_svd is slow
2026-01-14 09:54:49 -05:00
qazal
434dbafab5
optional Estimates in KernelInfo ( #14147 )
...
* optional Estimates in KernelInfo
* custom asm test plumbing
* s_code_end
* estimates test
* vaddr arg in global_store
* kernel desc
* Ops.DEVICE name
2026-01-14 22:55:03 +09:00
qazal
76b577ee76
viz: only SIMD name in sqtt timeline rows ( #14146 )
2026-01-14 20:13:27 +09:00
nimlgen
86708ccac5
hip_ioctl: dump aql ( #14142 )
2026-01-14 13:15:10 +03:00
wozeparrot
a92778aa0c
tk: fa multi fix ( #14134 )
2026-01-13 17:22:15 -08:00
George Hotz
2ab18ea7e3
assembly/amd: use xml instead of pdf ( #14118 )
...
* assembly/amd: use xml instead of pdf
* use amdxml to generate info about op sizes
* fix many tests with invalid instructions
* fix info generation
* chad xml fixes many bugs
* rename to operands
* simplify
* amdxml
* bug fix
2026-01-14 10:03:37 +09:00
qazal
002ea39da7
assembly/amd: use Tensor.custom_kernel to run assembly ( #14125 )
...
* assembly/amd: use Tensor.custom_kernel to run assembly
* PRINT_ASM=1 is DEBUG=4
2026-01-14 08:29:25 +09:00
chenyu
fe00682502
clean up svd tests ( #14133 )
...
removed from test_ops and added to TestTorchBackend
2026-01-13 16:32:21 -05:00
chenyu
e610821c52
Tensor.cummin and Tensor.nonzero ( #14131 )
2026-01-13 15:09:56 -05:00
chenyu
176a934ddd
Tensor.diagonal support offset and dims ( #14130 )
2026-01-13 14:49:06 -05:00
chenyu
2a217ba206
tinybackend isin and log10 ( #14120 )
...
can use tinygrad directly
2026-01-13 14:14:09 -05:00
George Hotz
a28c8105a5
assembly/amd: 2% faster amd_uop_matmul + SQTT ( #14122 )
...
* assembly/amd: 2% faster amd_uop_matmul
* SQTT_TOKEN_EXCLUDE + SQTT_SIMD_SEL
* sqtt printer
* fix printer
* fast decode
* fast decoder
* test packet counts
* ugh it's not faster
* dead
2026-01-13 19:55:32 +09:00
chenyu
05fcb57696
also return index in Tensor.cummax ( #14117 )
...
* also return index in Tensor.cummax
* fix
2026-01-12 22:42:10 -05:00
George Hotz
330a0b686e
assembly/amd: clean up dsl and make type verification strict ( #14102 )
...
* assembly/amd: start newdsl
* work
* newdsl upd
* Reg is p nice
* cleaner
* work
* getting clean
* all fields
* more BitFields
* redo the pdfs with dsl2 syntax
* no lit
* cleanups
* more defaults
* fix get and remove crap
* aliases
* ugly but kind of works
* NULL, not rawimm
* clean up defaults
* only dsl
* asm fixes
* lit fixup
* more lit
* cleanups
* olddsl
* single pcode dict
* emu sort of works
* trash test
* global is global
* types property
* reg mods
* fix a few tests
* remove monkey patch
* fixes
* less hacks in tests
* less hacks in tests
* 4 test failures
* hw tests all pass
* fix compare emulator
* fix some tests
* 3 more
* fix and shorten sqtt
* handwritten
* fix validation
* test corrections
* all types validate
* fix dsl2 tests
* fix bugs in disasm
* skips on cdna
* work
* repr with reg[]
* fix bitfield tests
* merge pcodes in dsl
* remove override
* disasm uses inst.types
* simpler
2026-01-13 08:52:16 +09:00
George Hotz
91bde927ef
assembly/amd: split asm.py into asm.py and disasm.py ( #14101 )
...
* split asm.py into asm.py and disasm.py
* split decoder
* move to pcode
* tests
2026-01-12 07:22:02 +09:00
George Hotz
44135e2e84
assembly/amd: always use v_nop in test for rocprof-trace-decoder ( #14100 )
...
* assembly/amd: always use v_nop in test for rocprof-trace-decoder
* test touchups
2026-01-12 05:31:58 +09:00
George Hotz
8b1b15aec0
assembly/amd: SQTT support ( #14099 )
...
* assembly/amd: SQTT support
* simpler
* cmp wave
* instruction compare
* rocprof decode
* simpler
* no llvm
* no strcmp
2026-01-12 05:07:17 +09:00
nimlgen
3e2c05ee9f
hevc: decoder as iterator ( #14091 )
2026-01-10 14:57:56 +03:00
b1tg
0fbc551622
train bert with fp8 ( #13874 )
...
* fp8 train
* clean
* lint
* test fix from #13439
* skip first/last layer
* rm __init__, restore unroll <=32 check
* tests
* clean test, remove unused
* multi-gpu test, clean quantize_to_fp8
* remove bert contiguous
* run script
* test: better check
* run script search
* add seed in bert data shuffle
* move script to mi350x folder
---------
Co-authored-by: chenyu <chenyu@fastmail.com >
2026-01-09 09:21:59 -05:00
qazal
2917ed1616
roc: propagate decoder errors to main thread ( #14081 )
...
* roc: propagate decoder errors to main thread
* types
* add cause
2026-01-09 21:10:45 +09:00
George Hotz
e7b5d8a434
assembly/amd: more RDNA4 asm ( #14062 )
...
* rdna4 more
* asm
* fixes
* assembly/amd: handwritten wmma failing test
* passes
* wmma default hacks
* space
* 0 skips in rdna3/rdna4 disasm
* more RDNA4 tests
---------
Co-authored-by: qazal <qazal.software@gmail.com >
2026-01-08 05:09:37 -08:00
nimlgen
e372c841ba
hevc: beam in decode ( #14067 )
...
* hevc: beam in decode
* fine
* g
2026-01-08 15:47:16 +03:00
nimlgen
f3aceaa08b
hevc: fast decoder ( #14057 )
2026-01-08 15:20:37 +03:00
qazal
309197bca5
assembly/amd: test_roundtrip for cdna/rdna4 ( #14066 )
2026-01-08 21:03:13 +09:00
qazal
15a056715d
fix amd assembly IDE tests on macbook ( #14063 )
2026-01-08 17:27:52 +09:00
