mirror of
https://github.com/tinygrad/tinygrad.git
synced 2026-04-29 03:00:14 -04:00
Merge branch 'master' into retinanet_mlperf
This commit is contained in:
Francis Lata
8
test/external/external_test_nv.py
vendored
8
test/external/external_test_nv.py
vendored
@@ -9,7 +9,7 @@ from tinygrad.engine.realize import get_runner, CompiledRunner
|
||||
from test.external.fuzz_linearizer import get_fuzz_rawbufs
|
||||
|
||||
from tinygrad.codegen.kernel import Kernel
|
||||
from tinygrad.ops import LazyOp, BinaryOps, UnaryOps, ReduceOps, BufferOps, MemBuffer
|
||||
from tinygrad.ops import LazyOp, Ops, ReduceOps, BufferOps, MemBuffer
|
||||
from tinygrad.shape.shapetracker import ShapeTracker
|
||||
from tinygrad.shape.view import View
|
||||
|
||||
@@ -26,12 +26,12 @@ class TestNV(unittest.TestCase):
|
||||
TestNV.addr = struct.pack("QQ", TestNV.b.lazydata.buffer._buf.va_addr, TestNV.a.lazydata.buffer._buf.va_addr)
|
||||
|
||||
def test_oor_kernels(self):
|
||||
ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=UnaryOps.CAST, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=UnaryOps.CAST, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 256, 1, 512, 4, 16, 4, 16), strides=(0, 100352, 0, 196, 0, 14, 0, 1), offset=-15, mask=((0, 1), (0, 256), (0, 1), (0, 512), (0, 4), (1, 15), (0, 4), (1, 15)), contiguous=False), View(shape=(256, 1, 512, 7, 7, 512, 3, 3), strides=(2097152, 0, 0, 128, 2, 4096, 1088, 17), offset=0, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(256, 1, 512, 7, 7, 512, 3, 3), strides=(25088, 0, 49, 7, 1, 0, 0, 0), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(dtypes.float, False)),), arg=((0, 3, 4), dtypes.float)),), arg=(dtypes.half, False)),), arg=MemBuffer(idx=0, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 1, 512, 1, 1, 512, 3, 3), strides=(0, 0, 4608, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=True),)))) # noqa: E501
|
||||
ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=Ops.CAST, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=Ops.CAST, src=(LazyOp(op=Ops.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 256, 1, 512, 4, 16, 4, 16), strides=(0, 100352, 0, 196, 0, 14, 0, 1), offset=-15, mask=((0, 1), (0, 256), (0, 1), (0, 512), (0, 4), (1, 15), (0, 4), (1, 15)), contiguous=False), View(shape=(256, 1, 512, 7, 7, 512, 3, 3), strides=(2097152, 0, 0, 128, 2, 4096, 1088, 17), offset=0, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(256, 1, 512, 7, 7, 512, 3, 3), strides=(25088, 0, 49, 7, 1, 0, 0, 0), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(dtypes.float, False)),), arg=((0, 3, 4), dtypes.float)),), arg=(dtypes.half, False)),), arg=MemBuffer(idx=0, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 1, 512, 1, 1, 512, 3, 3), strides=(0, 0, 4608, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=True),)))) # noqa: E501
|
||||
opts = [Opt(op=OptOps.TC, axis=6, amt=2), Opt(op=OptOps.UPCAST, axis=0, amt=4), Opt(op=OptOps.UPCAST, axis=3, amt=0), Opt(op=OptOps.LOCAL, axis=1, amt=4), Opt(op=OptOps.LOCAL, axis=2, amt=3), Opt(op=OptOps.UPCAST, axis=1, amt=2)] # noqa: E501
|
||||
helper_test_lin(Kernel(ast), opts=opts, failed_platforms=["NV"])
|
||||
|
||||
def test_error_on_huge_dims(self):
|
||||
ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=UnaryOps.CAST, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 1, 1024, 683), strides=(0, 0, 0, 1), offset=0, mask=None, contiguous=False),)))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 1, 1024, 683), strides=(0, 0, 683, 1), offset=0, mask=None, contiguous=True),))))), arg=None),), arg=dtypes.float),), arg=(3,)),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 1, 1024, 1), strides=(0, 0, 1, 0), offset=0, mask=None, contiguous=True),)))) # noqa: E501
|
||||
ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=Ops.CAST, src=(LazyOp(op=Ops.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 1, 1024, 683), strides=(0, 0, 0, 1), offset=0, mask=None, contiguous=False),)))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.half, st=ShapeTracker(views=(View(shape=(1, 1, 1024, 683), strides=(0, 0, 683, 1), offset=0, mask=None, contiguous=True),))))), arg=None),), arg=dtypes.float),), arg=(3,)),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 1, 1024, 1), strides=(0, 0, 1, 0), offset=0, mask=None, contiguous=True),)))) # noqa: E501
|
||||
opts = [Opt(op=OptOps.GROUP, axis=0, amt=0), Opt(op=OptOps.PADTO, axis=1, amt=32), Opt(op=OptOps.UNROLL, axis=0, amt=4), Opt(op=OptOps.LOCAL, axis=0, amt=2), Opt(op=OptOps.LOCAL, axis=0, amt=2)] # noqa: E501
|
||||
with self.assertRaises(RuntimeError) as cm:
|
||||
lin = Kernel(ast)
|
||||
@@ -43,7 +43,7 @@ class TestNV(unittest.TestCase):
|
||||
|
||||
def test_buf4_usage(self):
|
||||
TestNV.along = Tensor([105615], device="NV").realize()
|
||||
ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=UnaryOps.SIN, src=(LazyOp(op=UnaryOps.CAST, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.ulong, st=ShapeTracker(views=(View(shape=(3,), strides=(1,), offset=0, mask=None, contiguous=True),)))),), arg=dtypes.float),), arg=None),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(3,), strides=(1,), offset=0, mask=None, contiguous=True),)))) # noqa: E501
|
||||
ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=Ops.SIN, src=(LazyOp(op=Ops.CAST, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.ulong, st=ShapeTracker(views=(View(shape=(3,), strides=(1,), offset=0, mask=None, contiguous=True),)))),), arg=dtypes.float),), arg=None),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(3,), strides=(1,), offset=0, mask=None, contiguous=True),)))) # noqa: E501
|
||||
temp_runner = get_runner(TestNV.d0.dname, (ast,))
|
||||
temp_runner([TestNV.b.lazydata.buffer, TestNV.along.lazydata.buffer], var_vals={})
|
||||
val = TestNV.b.lazydata.buffer.as_buffer().cast("f")[0]
|
||||
|
||||
4
test/external/external_test_valid_remove.py
vendored
4
test/external/external_test_valid_remove.py
vendored
@@ -2,7 +2,7 @@
|
||||
import unittest
|
||||
|
||||
from tinygrad import Device
|
||||
from tinygrad.ops import UOp, Ops, BinaryOps
|
||||
from tinygrad.ops import UOp, Ops
|
||||
from tinygrad.engine.search import Opt, OptOps
|
||||
from tinygrad.dtype import dtypes
|
||||
from tinygrad.shape.shapetracker import ShapeTracker
|
||||
@@ -20,7 +20,7 @@ class TestOpenpilotValidhack(unittest.TestCase):
|
||||
UOp(Ops.ADD, dtypes.float, arg=None, src=(
|
||||
UOp(Ops.MAX, dtypes.float, arg=None, src=(
|
||||
x5:=UOp(Ops.ADD, dtypes.float, arg=None, src=(
|
||||
UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(BinaryOps.ADD, (7, 8, 9, 10)), src=(
|
||||
UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (7, 8, 9, 10)), src=(
|
||||
UOp(Ops.CAST, dtypes.float, arg=None, src=(
|
||||
UOp(Ops.MUL, dtypes.float, arg=None, src=(
|
||||
UOp(Ops.LOAD, dtypes.float, arg=None, src=(
|
||||
|
||||
4
test/external/fuzz_linearizer.py
vendored
4
test/external/fuzz_linearizer.py
vendored
@@ -25,7 +25,7 @@ from tinygrad.codegen.kernel import Opt, OptOps
|
||||
from tinygrad.engine.search import get_kernel_actions, bufs_from_lin
|
||||
from tinygrad.engine.realize import CompiledRunner
|
||||
from tinygrad.helpers import getenv, from_mv, prod, colored, Context, DEBUG, Timing
|
||||
from tinygrad.ops import UnaryOps, UOp, Ops
|
||||
from tinygrad.ops import UOp, Ops
|
||||
from tinygrad.device import is_dtype_supported
|
||||
|
||||
def on_linearizer_will_run(): pass
|
||||
@@ -252,7 +252,7 @@ def fuzz_linearizer(lin: Kernel, rtol=1e-2, atol=1e-2, opts_list=None):
|
||||
def _is_simple(lin: Kernel) -> bool:
|
||||
if len(lin.ast.src) > 1: return False
|
||||
ast:UOp = lin.ast.src[0]
|
||||
if ast.src[0].op is UnaryOps.CAST and ast.src[0].src[0].op is Ops.LOAD: return True
|
||||
if ast.src[0].op is Ops.CAST and ast.src[0].src[0].op is Ops.LOAD: return True
|
||||
return False
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
8
test/external/fuzz_schedule.py
vendored
8
test/external/fuzz_schedule.py
vendored
@@ -6,7 +6,7 @@ from tinygrad.engine.realize import capturing, lower_schedule_item
|
||||
from tinygrad.helpers import DEBUG, MULTIOUTPUT, colored, getenv
|
||||
from tinygrad.engine.lazy import LazyBuffer
|
||||
from tinygrad.engine.schedule import LBScheduleItem, _graph_schedule, ScheduleItem
|
||||
from tinygrad.ops import MetaOps
|
||||
from tinygrad.ops import Ops
|
||||
from tinygrad.tensor import Tensor, _to_np_dtype
|
||||
|
||||
ctx_vars = { MULTIOUTPUT: (0, 1) }
|
||||
@@ -33,7 +33,7 @@ def fuzz_schedule(outs:List[LazyBuffer]):
|
||||
for lsi in ts:
|
||||
for out in lsi.outputs:
|
||||
# freeze assign state before exec
|
||||
if out.op is MetaOps.ASSIGN:
|
||||
if out.op is Ops.ASSIGN:
|
||||
prerealized[out] = out.buffer.as_buffer()
|
||||
assign_targets[out.srcs[1]] = out
|
||||
for x in lsi.inputs:
|
||||
@@ -50,9 +50,9 @@ def fuzz_schedule(outs:List[LazyBuffer]):
|
||||
rawbufs: Dict[LazyBuffer, Buffer] = {}
|
||||
for lsi in ts:
|
||||
for out in lsi.outputs:
|
||||
base = rawbufs[lsi.inputs[0]].base if out.op is MetaOps.BUFFER_VIEW else None
|
||||
base = rawbufs[lsi.inputs[0]].base if out.op is Ops.BUFFER_VIEW else None
|
||||
rawbufs[out] = Buffer(out.buffer.device, out.buffer.size, out.buffer.dtype, base=base)
|
||||
if out.op is MetaOps.ASSIGN: rawbufs[out].ensure_allocated().copyin(prerealized[out])
|
||||
if out.op is Ops.ASSIGN: rawbufs[out].ensure_allocated().copyin(prerealized[out])
|
||||
for x in lsi.inputs:
|
||||
if x not in rawbufs:
|
||||
# override the assign_target after ASSIGN
|
||||
|
||||
6
test/external/fuzz_symbolic.py
vendored
6
test/external/fuzz_symbolic.py
vendored
@@ -42,9 +42,9 @@ def gt(expr, rng=None):
|
||||
return expr > rng, rng
|
||||
|
||||
# NOTE: you have to replace these for this test to pass
|
||||
from tinygrad.ops import python_alu, BinaryOps
|
||||
python_alu[BinaryOps.MOD] = lambda x,y: x%y
|
||||
python_alu[BinaryOps.IDIV] = lambda x,y: x//y
|
||||
from tinygrad.ops import python_alu, Ops
|
||||
python_alu[Ops.MOD] = lambda x,y: x%y
|
||||
python_alu[Ops.IDIV] = lambda x,y: x//y
|
||||
|
||||
if __name__ == "__main__":
|
||||
ops = [add_v, div, mul, add_num, mod]
|
||||
|
||||
Reference in New Issue
Block a user