mawei/StarPilot
1
0
Fork 0
mirror of https://github.com/firestar5683/StarPilot.git synced 2026-07-20 00:32:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

Tinygrad

Browse Source
This commit is contained in:
firestar5683
2025-09-30 08:56:38 -05:00
parent 255c89c6ab
commit aa9364a311
543 changed files with 270571 additions and 17904 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tinygrad_repo/tinygrad/runtime/ops_python.py
+60 -70
View File
@@ -2,41 +2,31 @@
# a python uops emulator
# works to test the tensor cores, and all the uops in general
# this is the (living) definition of uops
from typing import Any, TYPE_CHECKING
from typing import Optional, Any, TYPE_CHECKING
import pickle, base64, itertools, time, struct, sys
from tinygrad.dtype import DType, dtypes, ImageDType, PtrDType, truncate, float_to_bf16
from tinygrad.dtype import DType, dtypes, ImageDType, PtrDType, truncate
from tinygrad.helpers import all_same, getenv, flatten, get_single_element
from tinygrad.device import Compiled, Compiler, Allocator
from tinygrad.codegen.opt import tc
from tinygrad.uop.ops import exec_alu, python_alu, Ops, UOp, GroupOp
from tinygrad.uop.ops import exec_alu, Ops, UOp, GroupOp
from tinygrad.renderer import Renderer
from tinygrad.renderer.cstyle import CUDARenderer, MetalRenderer, AMDRenderer, IntelRenderer, ClangRenderer
def storage_fmt_for_dtype(dtype: DType): return 'H' if dtype == dtypes.bfloat16 else dtype.fmt
def to_storage_scalar(x, dtype: DType):
if dtype == dtypes.bfloat16: return (struct.unpack('I', struct.pack('f', float_to_bf16(x)))[0] >> 16) & 0xFFFF
return x
def from_storage_scalar(x, dtype: DType):
if dtype == dtypes.bfloat16: return struct.unpack('f', struct.pack('I', (x & 0xFFFF) << 16))[0]
return x
def _load(m, i, dtype: DType):
def _load(m, i):
if i is None: return 0.0
if i < 0 or i >= len(m): raise IndexError(f"load out of bounds, size is {len(m)} and access is {i}")
return from_storage_scalar(m[i], dtype)
return m[i]
def load(inp, j, dtype: DType):
if len(inp) == 2: return [_load(m, x+j if x is not None else None, dtype) if gate else default for (m,x,gate),default in zip(*inp)]
return [_load(m, x+j if x is not None else None, dtype) for m,x,_ in inp[0]]
def load(inp, j=0):
if len(inp) == 2: return [_load(m, x+j if x is not None else None) if gate else default for (m,x,gate),default in zip(*inp)]
return [_load(m, x+j if x is not None else None) for m,x,_ in inp[0]]
def _store(m, i, v, dtype: DType):
def _store(m, i, v):
if i < 0 or i >= len(m): raise IndexError(f"store out of bounds, size is {len(m)}, access is {i}, value is {v}")
m[i] = to_storage_scalar(v, dtype)
m[i] = v
class PythonProgram:
def __init__(self, name:str, lib:bytes):
self.uops: list[tuple[Ops, DType|None, list[int], Any]] = pickle.loads(lib)
self.uops: list[tuple[Ops, Optional[DType], list[int], Any]] = pickle.loads(lib)
def __call__(self, *bufs, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1), vals:tuple[int, ...]=(), wait=False):
st = time.perf_counter()
warp = list(itertools.product(*[range(x) for x in local_size[::-1]]))
@@ -50,43 +40,45 @@ class PythonProgram:
loop_ends: dict[int, int] = {}
while i < len(self.uops):
uop, dtype, idp, arg = self.uops[i]
void_ops = {Ops.ENDRANGE, Ops.BARRIER, Ops.IF, Ops.ENDIF, Ops.SINK, Ops.NOOP, Ops.STORE}
void_ops = {Ops.STORE, Ops.ENDRANGE, Ops.BARRIER, Ops.IF, Ops.ENDIF, Ops.SINK}
if uop is Ops.DEFINE_ACC: idp = [idp[0]]
inp = [ul[v] for v in idp if self.uops[v][0] not in void_ops]
dtp = [dl[v] for v in idp if self.uops[v][0] not in void_ops]
if getenv("TRACE"): print(i, uop, dtype, arg, inp, dtp)
if uop is Ops.STORE:
assert len(inp) == 2, "expected store is ([(memory, offset, gate)], [value])"
if dtp[1].count > 1:
for j,val in enumerate(inp[1]):
for (m,o,g),v in zip(inp[0], val):
if g: _store(m, o+j, v)
else:
for (m,o,g),v in zip(*inp):
if g: _store(m, o, v)
i += 1
continue
if uop is Ops.ENDRANGE:
loop_ends[idp[0]] = i
i = idp[0]
continue
if uop in (Ops.BARRIER, Ops.IF, Ops.ENDIF, Ops.SINK, Ops.NOOP):
if uop in (Ops.BARRIER, Ops.IF, Ops.ENDIF, Ops.SINK):
# in the python emulator, the warp is always in sync
i += 1
continue
assert dtype is not None, f"{uop} is missing a dtype"
dl[i] = dtype
if uop is Ops.STORE:
for j,val in enumerate(inp[1] if dtp[1].count > 1 else [inp[1]]):
for (m,o,g),v in zip(inp[0], val):
if g: _store(m, o+j, v, dtp[1].scalar())
i += 1
continue
if uop in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG}:
assert isinstance(dtype, PtrDType), dtype
storage_fmt = storage_fmt_for_dtype(dtype.base.scalar())
if storage_fmt is None: raise RuntimeError(f"{dtype=} is not supported")
if TYPE_CHECKING or sys.version_info < (3, 12): assert storage_fmt != "e"
if uop is Ops.DEFINE_REG:
# REGs are per thread
ul[i] = [memoryview(bytearray(dtype.size*dtype.itemsize)).cast(storage_fmt) for _ in range(warp_size)]
else:
buf = memoryview(bytearray(dtype.size*dtype.itemsize)) if uop is not Ops.DEFINE_GLOBAL else pbufs.pop(0)
ul[i] = [buf.cast(storage_fmt)] * warp_size
if uop in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL}:
assert dtype.fmt is not None and isinstance(dtype, PtrDType)
if TYPE_CHECKING or sys.version_info < (3, 12): assert dtype.fmt != "e"
buf = memoryview(bytearray(dtype.size*dtype.itemsize)) if uop is Ops.DEFINE_LOCAL else pbufs.pop(0)
ul[i] = [buf.cast(dtype.fmt)] * warp_size
elif uop is Ops.DEFINE_VAR:
ul[i] = [pvals.pop(0)] * warp_size
elif uop is Ops.SPECIAL:
if arg[0][0] == 'g': ul[i] = [idxs[2-int(arg[0][-1])]] * warp_size
elif arg[0][0] == 'l': ul[i] = [x[2-int(arg[0][-1])] for x in warp]
elif uop is Ops.CONST: ul[i] = [arg] * warp_size
elif uop is Ops.DEFINE_ACC:
ul[i] = [[inp[0][0][0]] * warp_size for _ in range(dtype.count)] if dtype.count > 1 else [inp[0][0]] * warp_size
elif uop is Ops.INDEX:
ret:list = []
if isinstance(dtp[0], ImageDType):
@@ -109,17 +101,19 @@ class PythonProgram:
continue
elif uop is Ops.VECTORIZE: ul[i] = inp
elif uop is Ops.BITCAST:
packed = struct.pack(str(warp_size) + storage_fmt_for_dtype(dtp[0].scalar()), *[to_storage_scalar(x, dtp[0].scalar()) for x in inp[0]])
ul[i] = list(struct.unpack(str(warp_size) + storage_fmt_for_dtype(dtype.scalar()), packed))
ul[i] = [from_storage_scalar(x, dtype.scalar()) for x in ul[i]]
assert dtp[0].fmt and dtype.fmt
pack_format, unpack_format = str(warp_size) + dtp[0].fmt, str(warp_size) + dtype.fmt
ul[i] = list(struct.unpack(unpack_format, struct.pack(pack_format, *inp[0])))
elif uop is Ops.CAST:
ul[i] = [truncate.get(dtype, lambda dt: dt)(dtypes.as_const(x, dtype)) for x in inp[0]]
elif uop is Ops.LOAD:
if dtype.count > 1:
ul[i] = [load([inp[i][j] if i != 0 and dtp[i].count > 1 else inp[i] for i in range(len(inp))], j, dtype.scalar()) \
for j in range(dtype.count)]
ul[i] = [load([inp[i][j] if i != 0 and dtp[i].count > 1 else inp[i] for i in range(len(inp))], j) for j in range(dtype.count)]
else:
ul[i] = load(inp, 0, dtype)
ul[i] = load(inp)
elif uop is Ops.ASSIGN:
for j in range(len(inp[0])): inp[0][j] = inp[1][j]
ul[i] = inp[0]
elif uop is Ops.GEP: ul[i] = inp[0][get_single_element(arg)]
elif uop is Ops.WMMA:
# here are the models for the WMMA instruction on the different hardware
@@ -136,27 +130,24 @@ class PythonProgram:
out[elem_idx][goff+lane_id] += sum(a_elem(inp[0], _k, c_j, goff) * b_elem(inp[1], c_i, _k, goff) for _k in range(K))
return out
first_src_dtype = self.uops[idp[0]][1]
assert isinstance(first_src_dtype, DType) # mypy
dims, dtype_in, device, threads = arg[1], first_src_dtype.scalar(), arg[4], arg[5]
# TODO: refactor these to a shared TensorCoreLayout in kernel.py
if device == "METAL":
if arg[4] == "METAL":
# A (2 elements on 32 threads): row major
def a_b_elem(x, i, j, goff): return x[(i%2)][goff+(i//2)%2+(j%4)*2+(i//4)*8+(j//4)*16]
# (i, j), C, D (2 elements on 32 threads): row major same as A/B
def c_map(lane, elem): return (elem + ((lane%2)*2) + ((lane//8)%2)*4, ((lane//2)%4) + (lane//16)*4)
ul[i] = wmma_helper(32, 8, 2, 2, 2, a_b_elem, a_b_elem, c_map)
elif device == "AMD" and threads == 64:
elif arg[4] == "AMD" and arg[5] == 64:
def a_elem(x, k, row, goff): return x[k%4][goff + (k//4)*16 + row]
def b_elem(x, col, k, goff): return a_elem(x, k, col, goff) # pylint: disable=arguments-out-of-order
def c_map(lane, elem): return (lane%16, (lane//16)*4 + elem)
ul[i] = wmma_helper(64, 16, 4, 4, 4, a_elem, b_elem, c_map)
elif device == "AMD" and len(inp[0]) == 8: # RDNA4
elif arg[4] == "AMD" and len(inp[0]) == 8: # RDNA4
def a_elem(x, k, row, goff): return x[k - [0, 4, 4, 8][k//4]][goff + row + [0, 16, 0, 16][k//4]]
def b_elem(x, col, k, goff): return a_elem(x, k, col, goff)
def c_map(lane, elem): return (lane%16, (lane//16)*8 + elem)
ul[i] = wmma_helper(32, 16, 8, 8, 8, a_elem, b_elem, c_map)
elif device == "AMD":
elif arg[4] == "AMD":
# A (16 elements on 32 threads): col major, lane 16-32 == lane 0-15
def a_elem(x, k, row, goff):
assert x[k][goff+row] == x[k][goff+row+16], "warp elements not duplicated properly across lanes"
@@ -165,27 +156,27 @@ class PythonProgram:
def b_elem(x, col, k, goff): return a_elem(x, k, col, goff) # pylint: disable=arguments-out-of-order
def c_map(lane, elem): return (lane%16, lane//16+elem*2) # (i, j), C, D (8 elements on 32 threads): row major
ul[i] = wmma_helper(32, 16, 16, 16, 8, a_elem, b_elem, c_map)
elif device == "CUDA":
elif arg[4] == "CUDA":
# (col, row) given (lane, elem) for C & D (4 elements on 32 threads); shared by all tc shapes with M=16 N=8
def c_map(lane, elem): return (elem%2 + (lane%4)*2, lane//4 + (elem//2)*8)
if dims == (8,16,16):
if arg[1] == (8,16,16):
def a_elem(x, k, row, goff): return x[k%2 + (row//8)*2 + (k//8)*4][goff + (k//2)%4 + (row%8)*4]
def b_elem(x, col, k, goff): return x[k%2 + (k//8)*2][goff + (k//2)%4 + col*4]
ul[i] = wmma_helper(32, 16, 8, 4, 4, a_elem, b_elem, c_map)
elif dims == (8,16,8) and dtype_in == dtypes.half:
elif arg[1] == (8,16,8) and arg[2] == dtypes.half:
def a_elem(x, k, row, goff): return x[k%2 + (row//8)*2][goff + k//2 + (row%8)*4]
def b_elem(x, col, k, goff): return x[k%2][goff + k//2 + col*4]
ul[i] = wmma_helper(32, 8, 4, 2, 4, a_elem, b_elem, c_map)
elif dims == (8,16,8) and dtype_in == dtypes.float:
elif arg[1] == (8,16,8) and arg[2] == dtypes.float:
def a_elem(x, k, row, goff): return x[(k//4)*2 + row//8][goff + k%4 + (row%8)*4]
def b_elem(x, col, k, goff): return x[k//4][goff + k%4 + col*4]
ul[i] = wmma_helper(32, 8, 4, 2, 4, a_elem, b_elem, c_map)
else: raise NotImplementedError(f"unimplemented tensor core {arg}")
elif device == "INTEL":
elif arg[4] == "INTEL":
# A (16 elements on 8 threads)
def a_elem(x, k, row, goff): return x[k%2+row*2][goff+k//2]
# B (16 elements on 8 threads)
@@ -193,14 +184,14 @@ class PythonProgram:
# C, D (8 elements on 8 threads)
def c_map(lane, elem): return (lane, elem)
ul[i] = wmma_helper(8, 16, 16, 16, 8, a_elem, b_elem, c_map)
elif device == "CPU":
elif arg[4] == "CPU":
def elem(x, col, row, _): return x[col+row][0] # k is always 0
def c_map(_, elem): return (elem%16, elem//16)
ul[i] = wmma_helper(1, 1, 16, 16, 256, elem, elem, c_map)
else: raise NotImplementedError(f"unimplemented tensor core {arg}")
elif uop in GroupOp.ALU:
assert all_same([len(x) for x in inp]), f"{[len(x) for x in inp]} doesn't match on {uop}"
assert all_same([dtype] + dtp) or uop in {*GroupOp.Comparison, Ops.WHERE}, f"dtype mismatch on {uop}"
assert all_same([dtype] + dtp) or uop in {Ops.CMPNE, Ops.CMPLT, Ops.WHERE}, f"dtype mismatch on {uop}"
ul[i] = [exec_alu(uop, dtype, p) for p in zip(*inp)]
assert i in ul, (uop, dtype, idp, arg)
i += 1
@@ -208,16 +199,15 @@ class PythonProgram:
class PythonRenderer(Renderer):
device = "PYTHON"
code_for_op = python_alu
def __init__(self):
if getenv("EMULATE_METAL"): self.device, self.tensor_cores = "METAL", tc.metal
if getenv("EMULATE_AMD"): self.device, self.tensor_cores = "AMD", tc.amd_rdna3
if getenv("EMULATE_AMD_MFMA"): self.device, self.tensor_cores = "AMD", tc.amd_cdna
if getenv("EMULATE_AMD_RDNA4"): self.device, self.tensor_cores = "AMD", tc.amd_rdna4
if getenv("EMULATE_CUDA"): self.device, self.tensor_cores = "CUDA", tc.cuda_sm80
if getenv("EMULATE_CUDA_SM75"): self.device, self.tensor_cores = "CUDA", tc.cuda_sm75
if getenv("EMULATE_INTEL"): self.device, self.suffix, self.tensor_cores = "INTEL", "INTEL", tc.intel
if getenv("EMULATE_AMX"): self.device, self.tensor_cores = "CPU", tc.amx
if getenv("EMULATE_METAL"): self.device, self.tensor_cores = "METAL", MetalRenderer.tensor_cores
if getenv("EMULATE_AMD"): self.device, self.tensor_cores = "AMD", AMDRenderer.tensor_cores
if getenv("EMULATE_AMD_MFMA"): self.device, self.tensor_cores = "AMD", AMDRenderer.tensor_cores_mfma
if getenv("EMULATE_AMD_RDNA4"): self.device, self.tensor_cores = "AMD", AMDRenderer.tensor_cores_rdna4
if getenv("EMULATE_CUDA"): self.device, self.tensor_cores = "CUDA", CUDARenderer.tc_sm80
if getenv("EMULATE_CUDA_SM75"): self.device, self.tensor_cores = "CUDA", CUDARenderer.tc_sm75
if getenv("EMULATE_INTEL"): self.device, self.suffix, self.tensor_cores = "INTEL", "INTEL", IntelRenderer.tensor_cores
if getenv("EMULATE_AMX"): self.device, self.tensor_cores = "CPU", ClangRenderer.tensor_cores
def render(self, uops:list[UOp]) -> str:
lops = [(u.op, u.dtype, [uops.index(v) for v in u.src], u.arg) for u in uops]