diff --git a/test/test_linearizer_overflows.py b/test/test_linearizer_overflows.py index 4d14d597..e4ec98b5 100644 --- a/test/test_linearizer_overflows.py +++ b/test/test_linearizer_overflows.py @@ -7,7 +7,8 @@ from tinygrad.engine.search import Opt, OptOps from tinygrad.engine.search import time_linearizer, bufs_from_lin # stuff needed to unpack a kernel -from extra.ops import LazyOp, BinaryOps, UnaryOps, ReduceOps, BufferOps, MemBuffer, ConstBuffer +from tinygrad.ops import UOp, UOps, BinaryOps, UnaryOps +from tinygrad.dtype import PtrDType from tinygrad.shape.shapetracker import ShapeTracker from tinygrad.shape.view import View @@ -23,43 +24,143 @@ def _test_overflow(ast, opts): @unittest.skip("unneeded without launch bounds") class TestLinearizerOverflow(unittest.TestCase): def test_overflow_1(self): - ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=BinaryOps.MAX, src=(LazyOp(op=BinaryOps.ADD, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BinaryOps.ADD, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 64, 1, 3, 8, 230, 8, 230), strides=(0, 150528, 0, 50176, 0, 224, 0, 1), offset=-675, mask=((0, 1), (0, 64), (0, 1), (0, 3), (0, 8), (3, 227), (0, 8), (3, 227)), contiguous=False), View(shape=(64, 1, 64, 112, 112, 3, 7, 7), strides=(10156800, 0, 0, 3680, 2, 3385600, 425040, 231), offset=0, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 3, 7, 7), strides=(0, 0, 147, 0, 0, 49, 7, 1), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(7, 6, 5)), LazyOp(op=BufferOps.CONST, src=(), arg=ConstBuffer(val=0.0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)))),), arg=None), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=3, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),))))), arg=None), LazyOp(op=UnaryOps.SQRT, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.CONST, src=(), arg=ConstBuffer(val=1.0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)))), LazyOp(op=UnaryOps.RECIP, src=(LazyOp(op=BinaryOps.ADD, src=(LazyOp(op=BufferOps.CONST, src=(), arg=ConstBuffer(val=1.0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)))), LazyOp(op=BufferOps.CONST, src=(), arg=ConstBuffer(val=1e-05, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),))))), arg=None),))), arg=None),), arg=None)), arg=None), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=4, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),))))), arg=None), LazyOp(op=BufferOps.CONST, src=(), arg=ConstBuffer(val=0.0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(802816, 0, 12544, 112, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)))) + ast = UOp(UOps.SINK, None, arg=None, src=( + UOp(UOps.STORE, None, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=0, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(802816, 0, 12544, 112, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()), + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MAX, src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.ADD, src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.ADD, src=( + UOp(UOps.REDUCE_AXIS, dtypes.float, arg=(BinaryOps.ADD, (7, 6, 5)), src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=1, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(1, 64, 1, 3, 8, 230, 8, 230), strides=(0, 150528, 0, 50176, 0, 224, 0, 1), offset=-675, mask=((0, 1), (0, 64), (0, 1), (0, 3), (0, 8), (3, 227), (0, 8), (3, 227)), contiguous=False), View(shape=(64, 1, 64, 112, 112, 3, 7, 7), strides=(10156800, 0, 0, 3680, 2, 3385600, 425040, 231), offset=0, mask=None, contiguous=False))), src=()),)), + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=2, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 3, 7, 7), strides=(0, 0, 147, 0, 0, 49, 7, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)), + x16:=UOp(UOps.CONST, dtypes.float, arg=0.0, src=( + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)), + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=3, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)), + UOp(UOps.ALU, dtypes.float, arg=UnaryOps.SQRT, src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + x23:=UOp(UOps.CONST, dtypes.float, arg=1.0, src=( + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)), + UOp(UOps.ALU, dtypes.float, arg=UnaryOps.RECIP, src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.ADD, src=( + x23, + UOp(UOps.CONST, dtypes.float, arg=1e-05, src=( + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 0, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)),)), + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=4, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(64, 1, 64, 112, 112, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)), + x16,)),)),)) opts = [Opt(op=OptOps.LOCAL, axis=3, amt=16), Opt(op=OptOps.LOCAL, axis=2, amt=16), Opt(op=OptOps.UPCAST, axis=1, amt=4), Opt(op=OptOps.UPCAST, axis=0, amt=4), Opt(op=OptOps.UPCAST, axis=2, amt=0)] _test_overflow(ast, opts) # From BEAM on hlb_cifar.py def test_overflow_2(self): - ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 512, 1, 32, 4, 34, 4, 34), strides=(0, 32768, 0, 1024, 0, 32, 0, 1), offset=-33, mask=((0, 1), (0, 512), (0, 1), (0, 32), (0, 4), (1, 33), (0, 4), (1, 33)), contiguous=False), View(shape=(512, 1, 64, 32, 32, 32, 3, 3), strides=(591872, 0, 0, 136, 1, 18496, 4760, 35), offset=0, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(512, 1, 64, 32, 32, 32, 3, 3), strides=(0, 0, 288, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(7, 6, 5)),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(512, 1, 64, 32, 32, 1, 1, 1), strides=(65536, 0, 1024, 32, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)))) + ast = UOp(UOps.SINK, None, arg=None, src=( + UOp(UOps.STORE, None, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=0, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(512, 1, 64, 32, 32, 1, 1, 1), strides=(65536, 0, 1024, 32, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()), + UOp(UOps.REDUCE_AXIS, dtypes.float, arg=(BinaryOps.ADD, (7, 6, 5)), src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=1, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(1, 512, 1, 32, 4, 34, 4, 34), strides=(0, 32768, 0, 1024, 0, 32, 0, 1), offset=-33, mask=((0, 1), (0, 512), (0, 1), (0, 32), (0, 4), (1, 33), (0, 4), (1, 33)), contiguous=False), View(shape=(512, 1, 64, 32, 32, 32, 3, 3), strides=(591872, 0, 0, 136, 1, 18496, 4760, 35), offset=0, mask=None, contiguous=False))), src=()),)), + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=2, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(512, 1, 64, 32, 32, 32, 3, 3), strides=(0, 0, 288, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)) opts = [Opt(op=OptOps.LOCAL, axis=3, amt=16), Opt(op=OptOps.LOCAL, axis=2, amt=4), Opt(op=OptOps.UPCAST, axis=1, amt=4), Opt(op=OptOps.UPCAST, axis=2, amt=0), Opt(op=OptOps.UPCAST, axis=1, amt=4), Opt(op=OptOps.UNROLL, axis=0, amt=0)] _test_overflow(ast, opts) # from BEAM on default simple_conv.py (which is quite large): def test_overflow_3(self): - ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 16, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 16), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(16, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(16, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(7, 6, 5)),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(16, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)))) + ast = UOp(UOps.SINK, None, arg=None, src=( + UOp(UOps.STORE, None, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=0, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(16, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()), + UOp(UOps.REDUCE_AXIS, dtypes.float, arg=(BinaryOps.ADD, (7, 6, 5)), src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=1, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(1, 16, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 16), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(16, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))), src=()),)), + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=2, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(16, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)) opts = [Opt(op=OptOps.LOCAL, axis=3, amt=16), Opt(op=OptOps.LOCAL, axis=2, amt=8), Opt(op=OptOps.UPCAST, axis=1, amt=4), Opt(op=OptOps.UPCAST, axis=3, amt=0), Opt(op=OptOps.UPCAST, axis=1, amt=2), Opt(op=OptOps.UPCAST, axis=2, amt=2)] _test_overflow(ast, opts) # from BEAM on BS=4 simple_conv.py: def test_overflow_4(self): - ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 4, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 4), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(4, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(4, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(7, 6, 5)),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(4, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)))) + ast = UOp(UOps.SINK, None, arg=None, src=( + UOp(UOps.STORE, None, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=0, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(4, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()), + UOp(UOps.REDUCE_AXIS, dtypes.float, arg=(BinaryOps.ADD, (7, 6, 5)), src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=1, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(1, 4, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 4), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(4, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))), src=()),)), + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=2, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(4, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)) opts = [Opt(op=OptOps.UPCAST, axis=3, amt=4), Opt(op=OptOps.LOCAL, axis=3, amt=16), Opt(op=OptOps.UPCAST, axis=1, amt=4), Opt(op=OptOps.LOCAL, axis=2, amt=4), Opt(op=OptOps.UPCAST, axis=1, amt=2), Opt(op=OptOps.UPCAST, axis=2, amt=4)] _test_overflow(ast, opts) # from BEAM on BS=2 simple_conv.py: def test_overflow_5(self): - ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 2, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 2), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(2, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(2, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(7, 6, 5)),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(2, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)))) + ast = UOp(UOps.SINK, None, arg=None, src=( + UOp(UOps.STORE, None, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=0, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(2, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()), + UOp(UOps.REDUCE_AXIS, dtypes.float, arg=(BinaryOps.ADD, (7, 6, 5)), src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=1, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(1, 2, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 2), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(2, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))), src=()),)), + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=2, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(2, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)) opts = [Opt(op=OptOps.LOCAL, axis=3, amt=16), Opt(op=OptOps.UPCAST, axis=1, amt=4), Opt(op=OptOps.UPCAST, axis=3, amt=0), Opt(op=OptOps.LOCAL, axis=2, amt=2), Opt(op=OptOps.UPCAST, axis=1, amt=2), Opt(op=OptOps.UPCAST, axis=2, amt=2)] _test_overflow(ast, opts) # from BEAM on BS=3 simple_conv.py: def test_overflow_6(self): - ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 3, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 3), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(3, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(3, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(7, 6, 5)),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(3, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)))) + ast = UOp(UOps.SINK, None, arg=None, src=( + UOp(UOps.STORE, None, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=0, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(3, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()), + UOp(UOps.REDUCE_AXIS, dtypes.float, arg=(BinaryOps.ADD, (7, 6, 5)), src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=1, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(1, 3, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 3), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(3, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))), src=()),)), + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=2, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(3, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)) opts = [Opt(op=OptOps.LOCAL, axis=3, amt=16), Opt(op=OptOps.UPCAST, axis=3, amt=0), Opt(op=OptOps.UPCAST, axis=1, amt=4), Opt(op=OptOps.LOCAL, axis=2, amt=8), Opt(op=OptOps.UPCAST, axis=1, amt=2), Opt(op=OptOps.UPCAST, axis=3, amt=2)] _test_overflow(ast, opts) # from BEAM on BS=3 simple_conv.py: (alt) def test_overflow_7(self): - ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 3, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 3), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(3, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(3, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(7, 6, 5)),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(3, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)))) + ast = UOp(UOps.SINK, None, arg=None, src=( + UOp(UOps.STORE, None, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=0, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(3, 1, 128, 128, 128, 1, 1, 1), strides=(2097152, 0, 16384, 128, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()), + UOp(UOps.REDUCE_AXIS, dtypes.float, arg=(BinaryOps.ADD, (7, 6, 5)), src=( + UOp(UOps.ALU, dtypes.float, arg=BinaryOps.MUL, src=( + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=1, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(1, 3, 1, 128, 4, 130, 4, 130), strides=(0, 2097152, 0, 16384, 0, 128, 0, 1), offset=-129, mask=((0, 1), (0, 3), (0, 1), (0, 128), (0, 4), (1, 129), (0, 4), (1, 129)), contiguous=False), View(shape=(3, 1, 128, 128, 128, 128, 3, 3), strides=(34611200, 0, 0, 520, 1, 270400, 68120, 131), offset=0, mask=None, contiguous=False))), src=()),)), + UOp(UOps.LOAD, dtypes.float, arg=None, src=( + UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=2, src=()), + UOp(UOps.SHAPETRACKER, None, arg=ShapeTracker(views=(View(shape=(3, 1, 128, 128, 128, 128, 3, 3), strides=(0, 0, 1152, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)) opts = [Opt(op=OptOps.UPCAST, axis=3, amt=4), Opt(op=OptOps.LOCAL, axis=3, amt=16), Opt(op=OptOps.UPCAST, axis=1, amt=4), Opt(op=OptOps.LOCAL, axis=2, amt=8), Opt(op=OptOps.UPCAST, axis=1, amt=2), Opt(op=OptOps.UPCAST, axis=2, amt=4)] _test_overflow(ast, opts) @@ -68,20 +169,26 @@ class TestLinearizerOverflow(unittest.TestCase): class TestLinearizerOverflowAlt(unittest.TestCase): def test_overflow_1(self): BS = 2 - in_1 = MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, BS, 1, 3, 8, 230, 8, 230), strides=(0, 150528, 0, 50176, 0, 224, 0, 1), offset=-675, mask=((0, 1), (0, BS), (0, 1), (0, 3), (0, 8), (3, 227), (0, 8), (3, 227)), contiguous=False), - View(shape=(BS, 1, 64, 112, 112, 3, 7, 7), strides=(10156800, 0, 0, 3680, 2, 3385600, 425040, 231), offset=0, mask=None, contiguous=False)))) - in_2 = MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(BS, 1, 64, 112, 112, 3, 7, 7), strides=(0, 0, 147, 0, 0, 49, 7, 1), offset=0, mask=None, contiguous=False),))) - ot_0 = MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(BS, 1, 64, 112, 112, 1, 1, 1), strides=(802816, 0, 12544, 112, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),))) - ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, arg=in_1), LazyOp(op=BufferOps.LOAD, arg=in_2))),), arg=(7, 6, 5)),), arg=ot_0) + g0, g1, g2 = [UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=i) for i in range(3)] + in_st_1 = ShapeTracker(views=(View(shape=(1, BS, 1, 3, 8, 230, 8, 230), strides=(0, 150528, 0, 50176, 0, 224, 0, 1), offset=-675, mask=((0, 1), (0, BS), (0, 1), (0, 3), (0, 8), (3, 227), (0, 8), (3, 227)), contiguous=False), + View(shape=(BS, 1, 64, 112, 112, 3, 7, 7), strides=(10156800, 0, 0, 3680, 2, 3385600, 425040, 231), offset=0, mask=None, contiguous=False))).to_uop() + in_st_2 = ShapeTracker(views=(View(shape=(BS, 1, 64, 112, 112, 3, 7, 7), strides=(0, 0, 147, 0, 0, 49, 7, 1), offset=0, mask=None, contiguous=False),)).to_uop() + ot_st = ShapeTracker(views=(View(shape=(BS, 1, 64, 112, 112, 1, 1, 1), strides=(802816, 0, 12544, 112, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)).to_uop() + prod = UOp(UOps.LOAD, dtypes.float, (g1, in_st_1)) * UOp(UOps.LOAD, dtypes.float, (g2, in_st_2)) + store = UOp(UOps.STORE, src=(g0, ot_st, UOp(UOps.REDUCE_AXIS, dtypes.float, (prod,), (BinaryOps.ADD, (7, 6, 5))))) + ast = UOp(UOps.SINK, src=(store,)) opts = [Opt(op=OptOps.LOCAL, axis=3, amt=16), Opt(op=OptOps.LOCAL, axis=2, amt=2), Opt(op=OptOps.UPCAST, axis=0, amt=2)] _test_overflow(ast, opts) def test_overflow_2(self): BS = 2 - in_1 = MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, BS, 1, 3, 8, 230, 8, 230), strides=(0, 150528, 0, 50176, 0, 224, 0, 1), offset=-675, mask=((0, 1), (0, BS), (0, 1), (0, 3), (0, 8), (3, 227), (0, 8), (3, 227)), contiguous=False), - View(shape=(BS, 1, 64, 112, 112, 3, 7, 7), strides=(10156800, 0, 0, 3680, 2, 3385600, 425040, 231), offset=0, mask=None, contiguous=False)))) - in_2 = MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(BS, 1, 64, 112, 112, 3, 7, 7), strides=(0, 0, 147, 0, 0, 49, 7, 1), offset=0, mask=None, contiguous=False),))) - ot_0 = MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(BS, 1, 64, 112, 112, 1, 1, 1), strides=(802816, 0, 12544, 112, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),))) - ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.SUM, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BufferOps.LOAD, arg=in_1), LazyOp(op=BufferOps.LOAD, arg=in_2))),), arg=(7, 6, 5)),), arg=ot_0) + g0, g1, g2 = [UOp(UOps.DEFINE_GLOBAL, PtrDType(dtypes.float), arg=i) for i in range(3)] + in_st_1 = ShapeTracker(views=(View(shape=(1, BS, 1, 3, 8, 230, 8, 230), strides=(0, 150528, 0, 50176, 0, 224, 0, 1), offset=-675, mask=((0, 1), (0, BS), (0, 1), (0, 3), (0, 8), (3, 227), (0, 8), (3, 227)), contiguous=False), + View(shape=(BS, 1, 64, 112, 112, 3, 7, 7), strides=(10156800, 0, 0, 3680, 2, 3385600, 425040, 231), offset=0, mask=None, contiguous=False))).to_uop() + in_st_2 = ShapeTracker(views=(View(shape=(BS, 1, 64, 112, 112, 3, 7, 7), strides=(0, 0, 147, 0, 0, 49, 7, 1), offset=0, mask=None, contiguous=False),)).to_uop() + ot_st = ShapeTracker(views=(View(shape=(BS, 1, 64, 112, 112, 1, 1, 1), strides=(802816, 0, 12544, 112, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)).to_uop() + prod = UOp(UOps.LOAD, dtypes.float, (g1, in_st_1)) * UOp(UOps.LOAD, dtypes.float, (g2, in_st_2)) + store = UOp(UOps.STORE, src=(g0, ot_st, UOp(UOps.REDUCE_AXIS, dtypes.float, (prod,), (BinaryOps.ADD, (7, 6, 5))))) + ast = UOp(UOps.SINK, src=(store,)) opts = [Opt(op=OptOps.LOCAL, axis=3, amt=16), Opt(op=OptOps.UPCAST, axis=1, amt=4), Opt(op=OptOps.LOCAL, axis=2, amt=16), Opt(op=OptOps.UPCAST, axis=4, amt=4), Opt(op=OptOps.UPCAST, axis=1, amt=2), Opt(op=OptOps.UPCAST, axis=5, amt=2)] _test_overflow(ast, opts)