* most of the work from the uops2 branch
* schedule
* realize
* kernel
* lowerer
* search
* green
* merge uops with ops
* Revert "merge uops with ops"
This reverts commit 1408a59f12.
* fix benchmark
* remove extra dedup
* fixed xmx demo
* i think i'm invoking the DPAS but it's slow
* compiler build arg to stop register spilling, indicated where to fix flop counter
* don't mind this
* do NOT mind me
* do not mind me
* do not view
* i will add bf16 later
* in process of figuring out tc fields
* we figured out the fields!!!
* added check for cl device vendor, added seperate IntelRenderer
* remove tc thread_local_aliases
* cleaning debris before draft pr
* edits for linter
* deduping and checking device extensions
* i will find more line reductions in other places
* before merge upstream
* double grf size in compiler to fix register spilling (bandaid), device checking changes
* tc python emulation
* fixed emulation
* tests for emulated intel tensor core
* TC=0, 1 working on upstream, fixed perf
* test
* debris
* check for specialized cl device when we canonicalize device
* bf16 support, tc=3 test added
* address tests
* revert half2 loads on intel tc, cleanup
* linter
* fold_expanded revert
* lint, whitespace fix
* cuda bf16 (only one with bf16) is skipped in test tensor cores, so i will skip for intel bf16 too
* make line shorter, no need for noqa E501
* removed device intel
* fix python emulation
---------
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>
- remove unused code
- add CP_REG_TO_MEM opcode
- fixed parse_cmd_buf for more than 1 command object by correcting
an offset
- fixed memory mappings for cases when memory was allocated with
KGSL_MEMFLAGS_USE_CPU_MAP.
KGSL_MEMFLAGS_USE_CPU_MAP: If set on call and return, the returned GPU
address will be 0. Calling mmap() will set the GPU address.
So there are no IOCTL_KGSL_GPUOBJ_INFO ioctls for that type of memory
and it resulted to crash right after get_mem.
* mcts graph and dedup support
* usable graph
* mcts colors
* C=4 seems better
* C=3 even better
* sample_tree
* backprop is external function
* late expand to match algo
MCTS 500 is competitive with BEAM=8 on resnet on M1 Max.
- increment trial times even with compiled error and runtime error.
- use best time of children as the node value.
* mcts search
* mcts cleanups
* mcts cleanup
* random shuffle children order
* mcts in handcode_opt
* src and remove_node
* debug 3 to print ast
* print the type
* mcts in extra
* test/external/fuzz_linearizer: fix for new AST changes
also add beautiful_mnist failures
* add CLANG and LLVM to test_failure_35 failed_platforms
* fix test_linearizer_failure names
* added model impl
* minor cleanups
* extracted weights loading into from_pretrained
* reorganized model for better weight loading
* removed lru cache for state dict loading
