forked from IQ.Lvbs/IQ.Pilot
IQ.Pilot Prebuilt Release @ ab07000
This commit is contained in:
IQ.Lvbs history cleanup
0
tinygrad_repo/test/opt/__init__.py
Normal file
0
tinygrad_repo/test/opt/__init__.py
Normal file
161
tinygrad_repo/test/opt/test_gen_float4.py
Normal file
161
tinygrad_repo/test/opt/test_gen_float4.py
Normal file
@@ -0,0 +1,161 @@
|
||||
import unittest
|
||||
from tinygrad import Device, Tensor, dtypes
|
||||
from tinygrad.uop.ops import UOp, Ops
|
||||
from tinygrad.codegen import to_program
|
||||
from tinygrad.codegen.opt import Opt, OptOps
|
||||
|
||||
from tinygrad.helpers import DEV
|
||||
from test.helpers import replace_opts
|
||||
|
||||
AMX = "AMX" in DEV.arch
|
||||
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "need backends that support float4")
|
||||
class TestFloat4(unittest.TestCase):
|
||||
@staticmethod
|
||||
def count_float4(uops: list[UOp], n=4):
|
||||
return (len([uop for uop in uops if uop.op is Ops.LOAD and uop.dtype == dtypes.float.vec(n)]),
|
||||
len([uop for uop in uops if uop.op is Ops.STORE and uop.src[1].dtype == dtypes.float.vec(n)]))
|
||||
@staticmethod
|
||||
def count_half4(uops: list[UOp]):
|
||||
return (len([uop for uop in uops if uop.op is Ops.LOAD and uop.dtype == dtypes.half.vec(4)]),
|
||||
len([uop for uop in uops if uop.op is Ops.STORE and uop.src[1].dtype == dtypes.half.vec(4)]))
|
||||
|
||||
def test_float4_basic(self):
|
||||
a = Tensor.empty(2, 8).realize()
|
||||
b = Tensor.empty(2, 8).realize()
|
||||
c = a + b
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
realized_ast = s.src[0]
|
||||
opts_to_apply = [Opt(op=OptOps.UPCAST, axis=0, arg=4)]
|
||||
program = to_program(replace_opts(realized_ast, opts_to_apply), renderer=Device[Device.DEFAULT].renderer)
|
||||
|
||||
assert TestFloat4.count_float4(tuple(program.src[2].src)) == (2, 1)
|
||||
|
||||
@unittest.skipIf(Device.DEFAULT in {"CPU"} and AMX, "CPU with AMX upcasts float up to size 16")
|
||||
def test_float4_multidim(self):
|
||||
a = Tensor.empty(2, 8).realize()
|
||||
b = Tensor.empty(2, 8).realize()
|
||||
c = a + b
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
uops = tuple(to_program(replace_opts(s.src[0], [Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=2)]),
|
||||
renderer=Device[Device.DEFAULT].renderer).src[2].src)
|
||||
assert TestFloat4.count_float4(uops) == (4, 2)
|
||||
|
||||
@unittest.skipUnless(Device.DEFAULT in {"CPU"} and AMX, "Only CPU with AMX upcasts float up to size 16")
|
||||
def test_float4_multidim_amx(self):
|
||||
def kernel_for_shape(size, shift):
|
||||
a = Tensor.empty(2, size).realize()
|
||||
b = Tensor.empty(2, size).realize()
|
||||
c = a + b
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
return tuple(to_program(replace_opts(s.src[0], [Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=shift)]),
|
||||
renderer=Device[Device.DEFAULT].renderer).src[2].src)
|
||||
|
||||
sizes = [12, 8, 16]
|
||||
shifts = [3, 2, 4]
|
||||
expected_upcast_size = [4, 8, 16]
|
||||
expected_output = [(6,3), (2,1), (2,1)]
|
||||
|
||||
for i in range(len(sizes)):
|
||||
assert TestFloat4.count_float4(kernel_for_shape(sizes[i], shifts[i]), expected_upcast_size[i]) == expected_output[i]
|
||||
|
||||
def test_float4_unaligned_load(self):
|
||||
a = Tensor.empty(9).realize().shrink(((1, 9),))
|
||||
b = Tensor.empty(9).realize().shrink(((1, 9),))
|
||||
c = a + b
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
realized_ast = s.src[0]
|
||||
opts_to_apply = [Opt(op=OptOps.UPCAST, axis=0, arg=4)]
|
||||
program = to_program(replace_opts(realized_ast, opts_to_apply), renderer=Device[Device.DEFAULT].renderer)
|
||||
|
||||
assert TestFloat4.count_float4(tuple(program.src[2].src)) == (0, 1)
|
||||
|
||||
@unittest.skipIf(Device.DEFAULT in {"CPU"} and AMX, "CPU with AMX upcasts float up to size 16")
|
||||
def test_float4_multidim_unaligned_load(self):
|
||||
a = Tensor.empty(2, 9).realize().shrink(((0, 2), (1, 9),))
|
||||
b = Tensor.empty(2, 9).realize().shrink(((0, 2), (1, 9),))
|
||||
c = a + b
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
uops = tuple(to_program(replace_opts(s.src[0], [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=2)]),
|
||||
renderer=Device[Device.DEFAULT].renderer).src[2].src)
|
||||
|
||||
assert TestFloat4.count_float4(uops) == (0, 2)
|
||||
|
||||
@unittest.skipUnless(Device.DEFAULT in {"CPU"} and AMX, "Only CPU with AMX upcasts float up to size 16")
|
||||
def test_float4_multidim_unaligned_load_amx(self):
|
||||
def kernel_for_shape(size, shift):
|
||||
a = Tensor.empty(2, size).realize().shrink(((0, 2), (1, size),))
|
||||
b = Tensor.empty(2, size).realize().shrink(((0, 2), (1, size),))
|
||||
c = a + b
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
return tuple(to_program(replace_opts(s.src[0], [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=shift)]),
|
||||
renderer=Device[Device.DEFAULT].renderer).src[2].src)
|
||||
|
||||
sizes = [13, 9, 17]
|
||||
shifts = [3, 2, 4]
|
||||
expected_upcast_size = [4, 8, 16]
|
||||
expected_output = [(0,3), (0,1), (0,1)]
|
||||
|
||||
for i in range(len(sizes)):
|
||||
assert TestFloat4.count_float4(kernel_for_shape(sizes[i], shifts[i]), expected_upcast_size[i]) == expected_output[i]
|
||||
|
||||
def test_float4_sometimes_unaligned(self):
|
||||
a = Tensor.empty(1, 1, 8).realize()
|
||||
b = Tensor.empty(1, 1, 5).realize().shrink(((0, 1), (0, 1), (1, 5)))
|
||||
c = a.conv2d(b)
|
||||
# only the first and last conv dot products are aligned in a, and b is never aligned, so no
|
||||
# float4 should be emitted (the reduce axis of size 4 is the float4 axis here)
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
uops = tuple(to_program(replace_opts(s.src[0], [Opt(op=OptOps.UNROLL, axis=0, arg=4)]), renderer=Device[Device.DEFAULT].renderer).src[2].src)
|
||||
|
||||
assert TestFloat4.count_float4(uops) == (0, 0)
|
||||
|
||||
def test_float4_multidim_sometimes_unaligned(self):
|
||||
a = Tensor.empty(1, 1, 7).realize()
|
||||
b = Tensor.empty(1, 1, 5).realize().shrink(((0, 1), (0, 1), (1, 5)))
|
||||
c = a.conv2d(b)
|
||||
# the first conv dot product is aligned in a. If we upcast the output and reduce
|
||||
# dimension, then we could do float4 for only that one set of loads, but we currently
|
||||
# don't.
|
||||
# UPDATE: now we do this fusion
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
uops = tuple(to_program(replace_opts(s.src[0], [Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=0, arg=0)]),
|
||||
renderer=Device[Device.DEFAULT].renderer).src[2].src)
|
||||
|
||||
assert TestFloat4.count_float4(uops) in {(0,1), (1,1)}
|
||||
|
||||
def test_float4_expand(self):
|
||||
a = Tensor.empty(9).realize().shrink(((1, 9),))
|
||||
b = Tensor.empty(2).realize().reshape((2, 1)).expand((2,4)).reshape((8,))
|
||||
c = a + b
|
||||
|
||||
# we will upcast the top axis of sz 4. they should not be coalesced into float4,
|
||||
# since the top axis is not contiguous.
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
uops = tuple(to_program(replace_opts(s.src[0], [Opt(op=OptOps.UPCAST, axis=0, arg=4)]), renderer=Device[Device.DEFAULT].renderer).src[2].src)
|
||||
|
||||
assert TestFloat4.count_float4(uops) == (0, 1)
|
||||
|
||||
def test_float4_heterogeneous(self):
|
||||
a = Tensor.empty(8).realize()
|
||||
b = Tensor.empty(9).realize().shrink(((1, 9),))
|
||||
c = a + b
|
||||
|
||||
# should float4 b but not a
|
||||
|
||||
s = c.schedule_linear().src[0]
|
||||
uops = tuple(to_program(replace_opts(s.src[0], [Opt(op=OptOps.UPCAST, axis=0, arg=4)]), renderer=Device[Device.DEFAULT].renderer).src[2].src)
|
||||
|
||||
assert TestFloat4.count_float4(uops) == (1, 1)
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
364
tinygrad_repo/test/opt/test_kernel_opts.py
Normal file
364
tinygrad_repo/test/opt/test_kernel_opts.py
Normal file
@@ -0,0 +1,364 @@
|
||||
import unittest
|
||||
from tinygrad import Device, Tensor, dtypes
|
||||
from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
|
||||
|
||||
# TODO: write a clean version of this
|
||||
from test.backend.test_linearizer import helper_linearizer_opt
|
||||
|
||||
class TestKernelOpts(unittest.TestCase):
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
|
||||
def test_local_and_grouped_reduce(self):
|
||||
N = 128
|
||||
Tensor.manual_seed(1882)
|
||||
a = Tensor.rand(4, 4, N, N)
|
||||
b = Tensor.rand(4, 4, N)
|
||||
r = (b.sqrt() + ((a+1).sum(axis=3).exp()))
|
||||
helper_linearizer_opt(r, [
|
||||
[Opt(OptOps.LOCAL, 0, 2)],
|
||||
[Opt(OptOps.LOCAL, 0, 8)],
|
||||
[Opt(OptOps.LOCAL, 0, 16)], # Checking how it works with locals
|
||||
[Opt(OptOps.GROUPTOP, 0, 2)],
|
||||
[Opt(OptOps.GROUPTOP, 0, 32)],
|
||||
[Opt(OptOps.GROUPTOP, 0, 64)], # Checking how it works with grouped reduce
|
||||
[Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUPTOP, 0, 2)],
|
||||
[Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.GROUPTOP, 0, 16)],
|
||||
[Opt(OptOps.LOCAL, 0, 32), Opt(OptOps.GROUPTOP, 0, 2)],
|
||||
# Checking how it works with locals + grouped reduce
|
||||
[Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUPTOP, 0, 64)],
|
||||
# Checking how it works with locals + grouped reduce + upcasts
|
||||
[Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.UPCAST, 0, 8), Opt(OptOps.UNROLL, 1, 4)],
|
||||
# many local + many group
|
||||
[Opt(OptOps.GROUP, 0, 2)] * 4,
|
||||
[Opt(OptOps.LOCAL, 0, 2)] * 4,
|
||||
[Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUP, 0, 2)] * 4,
|
||||
])
|
||||
|
||||
def test_upcasts(self):
|
||||
N = 16
|
||||
Tensor.manual_seed(1772)
|
||||
a = Tensor.rand(N, N)
|
||||
b = Tensor.rand(N, N)
|
||||
r = (a+b).sqrt() * ((a+1).exp())
|
||||
helper_linearizer_opt(r, [
|
||||
[Opt(OptOps.UPCAST, 0, 2)],
|
||||
[Opt(OptOps.UPCAST, 0, 4)],
|
||||
[Opt(OptOps.UPCAST, 0, 8)], # Checking how it works with upcasts
|
||||
])
|
||||
|
||||
def test_full_upcast(self):
|
||||
Tensor.manual_seed(1772)
|
||||
a = Tensor.rand(4)
|
||||
b = Tensor.rand(4)
|
||||
r = (a+b).sqrt() * ((a+1).exp())
|
||||
helper_linearizer_opt(r, [
|
||||
[Opt(OptOps.UPCAST, 0, 4)], # Checking how it works with upcasts
|
||||
])
|
||||
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
|
||||
def test_matmul(self):
|
||||
N = 128
|
||||
Tensor.manual_seed(1552)
|
||||
a = Tensor.rand(N, N)
|
||||
b = Tensor.rand(N, N)
|
||||
r = a@b
|
||||
helper_linearizer_opt(r, [
|
||||
[Opt(OptOps.UPCAST, 0, 2)],
|
||||
[Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4)], # Checking how it works with upcasts
|
||||
[Opt(OptOps.LOCAL, 0, 2)],
|
||||
[Opt(OptOps.LOCAL, 1, 32)],
|
||||
[Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 4)],
|
||||
[Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 32)],
|
||||
[Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.LOCAL, 1, 8)], # Checking how it works with locals
|
||||
[Opt(OptOps.GROUPTOP, 0, 2)],
|
||||
[Opt(OptOps.GROUPTOP, 0, 32)],
|
||||
[Opt(OptOps.GROUPTOP, 0, 32), Opt(OptOps.UNROLL, 0, 4)], # Checking how it works with grouped_reduce
|
||||
[Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.LOCAL, 1, 2), Opt(OptOps.GROUPTOP, 0, 32)],
|
||||
[Opt(OptOps.LOCAL, 0, 8), Opt(OptOps.GROUPTOP, 0, 32)],
|
||||
[Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 0, 8), Opt(OptOps.GROUPTOP, 0, 4)], # Checking how it works with local+grouped_reduce
|
||||
# Checking all together
|
||||
[Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 4),
|
||||
Opt(OptOps.UPCAST, 1, 2)],
|
||||
# Full global upcast + local
|
||||
[Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 8)],
|
||||
])
|
||||
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
|
||||
def test_double_reduce(self):
|
||||
N = 128
|
||||
Tensor.manual_seed(1552)
|
||||
a = Tensor.rand(8, N, 8, N)
|
||||
r = a.sum(axis=(1,3))
|
||||
helper_linearizer_opt(r, [
|
||||
# openCL / DEV=CL is 256 max threads
|
||||
[Opt(OptOps.GROUPTOP, 0, 2)], [Opt(OptOps.GROUPTOP, 0, 32)],
|
||||
[Opt(OptOps.GROUPTOP, 1, 2)], [Opt(OptOps.GROUPTOP, 1, 32)], # Checking how it works with 1 grouped_reduce.
|
||||
[Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)],
|
||||
[Opt(OptOps.GROUPTOP, 0, 16), Opt(OptOps.GROUPTOP, 1, 2)],
|
||||
[Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 64)], # Checking how it works with 2 grouped_reduces.
|
||||
[Opt(OptOps.GROUPTOP, 0, 16), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.UNROLL, 0, 4)],
|
||||
[Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 32), Opt(OptOps.UNROLL, 2, 4)], # Checking how it works with 2 grouped_reduces + upcasts.
|
||||
[Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 4), Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4)],
|
||||
# Checking how it works with 2 grouped_reduces + upcasts + locals.
|
||||
[Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 4), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 32), Opt(OptOps.UNROLL, 1, 4)],
|
||||
[Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.LOCAL, 1, 2), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UPCAST, 0, 2)],
|
||||
[Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.LOCAL, 1, 2), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UPCAST, 0, 2),
|
||||
Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4)], # Checking how it works with 2 grouped_reduces + upcasts + locals.
|
||||
[Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 4), Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UPCAST, 0, 2),
|
||||
Opt(OptOps.UPCAST, 0, 2)], # No globals
|
||||
])
|
||||
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
@unittest.skipUnless(any(tc.dtype_in == tc.dtype_out == dtypes.half for tc in Device[Device.DEFAULT].renderer.tensor_cores),
|
||||
"test requires tensor cores with accumulation in half") # testing with half suffices.
|
||||
def test_tensor_core_opts(self):
|
||||
N = 128
|
||||
Tensor.manual_seed(1552)
|
||||
a, b = Tensor.rand(N, N, dtype=dtypes.half), Tensor.rand(N, N, dtype=dtypes.half)
|
||||
r = a.matmul(b, dtype=dtypes.half)
|
||||
atol, rtol = 0.25, 0.01
|
||||
helper_linearizer_opt(r, [
|
||||
[],
|
||||
[Opt(OptOps.UPCAST, 0, 4)],
|
||||
[Opt(OptOps.UPCAST, 1, 4)],
|
||||
[Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4)], # check upcasts
|
||||
[Opt(OptOps.UNROLL, 0, 2)], # check unroll
|
||||
[Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 0, 2)], # check combo of unroll and local
|
||||
[Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 2)],
|
||||
[Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 4)],
|
||||
[Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UPCAST, 0, 4)], # check permutations
|
||||
[Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 0, 4)],
|
||||
[Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 1, 4)],
|
||||
[Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 0, 4)],
|
||||
], apply_tc=True, atol=atol, rtol=rtol)
|
||||
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
@unittest.skipUnless(any(tc.dtype_in == tc.dtype_out == dtypes.half for tc in Device[Device.DEFAULT].renderer.tensor_cores),
|
||||
"test requires tensor cores with accumulation in half") # testing with half suffices.
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
|
||||
def test_tensor_core_opts_locals(self):
|
||||
N = 128
|
||||
Tensor.manual_seed(1552)
|
||||
a, b = Tensor.rand(N, N, dtype=dtypes.half), Tensor.rand(N, N, dtype=dtypes.half)
|
||||
r = a.matmul(b, dtype=dtypes.half)
|
||||
atol, rtol = 0.25, 0.01
|
||||
helper_linearizer_opt(r, [
|
||||
[Opt(OptOps.UNROLL, 0, 0)], # check full unroll of reduce with locals
|
||||
[Opt(OptOps.LOCAL, 0, 4)], # check local
|
||||
[Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.LOCAL, 0, 2)],
|
||||
[Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 0, 4)],
|
||||
], apply_tc=True, atol=atol, rtol=rtol)
|
||||
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared memory")
|
||||
@unittest.skipUnless(any(tc.dtype_in == tc.dtype_out == dtypes.half for tc in Device[Device.DEFAULT].renderer.tensor_cores),
|
||||
"test requires tensor cores with accumulation in half") # testing with half suffices.
|
||||
# NOTE: the METAL test is broken, likely due to a compiler bug. passes on CI with -O0 and with default opt level locally on M3
|
||||
@unittest.skipIf(Device.DEFAULT == "METAL", "broken for METAL")
|
||||
@unittest.skip("feature was removed")
|
||||
def test_tensor_core_opts_group(self):
|
||||
N = 128
|
||||
Tensor.manual_seed(1552)
|
||||
a, b = Tensor.rand(N, N, dtype=dtypes.half), Tensor.rand(N, N, dtype=dtypes.half)
|
||||
r = a.matmul(b, dtype=dtypes.half)
|
||||
atol, rtol = 0.25, 0.01
|
||||
helper_linearizer_opt(r, [
|
||||
[Opt(OptOps.GROUP, 0, 2)],
|
||||
[Opt(OptOps.GROUPTOP, 0, 4)],
|
||||
[Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.GROUP, 0, 2)],
|
||||
[Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.GROUP, 0, 2)],
|
||||
[Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.GROUP, 0, 2)],
|
||||
[Opt(OptOps.UPCAST, 0, 2), Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUP, 0, 2)],
|
||||
[Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 1, 2)],
|
||||
], apply_tc=True, atol=atol, rtol=rtol)
|
||||
|
||||
def test_padto_matmul(self):
|
||||
N = 17
|
||||
Tensor.manual_seed(289)
|
||||
a = Tensor.rand(N, N)
|
||||
b = Tensor.rand(N, N)
|
||||
helper_linearizer_opt(a@b, [
|
||||
[Opt(OptOps.PADTO, 0, 32)],
|
||||
[Opt(OptOps.PADTO, 1, 32)],
|
||||
[Opt(OptOps.PADTO, 2, 32)],
|
||||
[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.PADTO, 1, 32)],
|
||||
[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.PADTO, 1, 32), Opt(OptOps.PADTO, 2, 32)],
|
||||
# can optimize further post PADTO
|
||||
[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.PADTO, 1, 32), Opt(OptOps.UPCAST, 0, 2), Opt(OptOps.UPCAST, 1, 2),],
|
||||
])
|
||||
|
||||
def test_padto_upcasted_not_ok(self):
|
||||
N = 4
|
||||
a = Tensor.rand(N, N)
|
||||
b = Tensor.rand(N, N)
|
||||
helper_linearizer_opt(a@b, [
|
||||
[Opt(OptOps.UPCAST, 0, 0)],
|
||||
[Opt(OptOps.UPCAST, 1, 0)],
|
||||
[Opt(OptOps.UNROLL, 0, 0)],
|
||||
[Opt(OptOps.PADTO, 0, 8)],
|
||||
[Opt(OptOps.PADTO, 1, 8)],
|
||||
[Opt(OptOps.PADTO, 2, 8)],
|
||||
])
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(a@b, [[Opt(OptOps.UPCAST, 0, 0), Opt(OptOps.PADTO, 1, 8)]])
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(a@b, [[Opt(OptOps.UPCAST, 1, 0), Opt(OptOps.PADTO, 1, 8)]])
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(a@b, [[Opt(OptOps.UNROLL, 0, 0), Opt(OptOps.PADTO, 2, 8)]])
|
||||
|
||||
def test_padto_sum_ok(self):
|
||||
N = 18
|
||||
# NOTE: this setup prevents 17 * 17 contiguous merged into one dimension
|
||||
a = Tensor.rand(N, N).realize().shrink(((0, 17), (0, 17))) * 100
|
||||
b = (Tensor.rand(N, N) < 0.5).realize().shrink(((0, 17), (0, 17)))
|
||||
|
||||
helper_linearizer_opt(a.sum(0), [
|
||||
[Opt(OptOps.PADTO, 0, 32)],
|
||||
[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
|
||||
])
|
||||
helper_linearizer_opt(a.sum(1), [
|
||||
[Opt(OptOps.PADTO, 0, 32)],
|
||||
[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
|
||||
])
|
||||
|
||||
# can pad sum reduce axis if there's no unsafe ops prior to sum
|
||||
for axis in (0, 1):
|
||||
helper_linearizer_opt(a.sum(), [[Opt(OptOps.PADTO, axis, 32)],])
|
||||
helper_linearizer_opt(a.sum(0), [[Opt(OptOps.PADTO, axis, 32)],])
|
||||
helper_linearizer_opt(b.sum(), [[Opt(OptOps.PADTO, axis, 32)],])
|
||||
helper_linearizer_opt(b.sum(0), [[Opt(OptOps.PADTO, axis, 32)],])
|
||||
helper_linearizer_opt(b.sum(dtype=dtypes.bool), [[Opt(OptOps.PADTO, axis, 32)],])
|
||||
# TODO: why?
|
||||
if Device.DEFAULT != "WEBGPU":
|
||||
helper_linearizer_opt(b.sum(0, dtype=dtypes.bool), [[Opt(OptOps.PADTO, axis, 32)],])
|
||||
helper_linearizer_opt(b.sum(1, dtype=dtypes.bool), [[Opt(OptOps.PADTO, axis, 32)],])
|
||||
|
||||
# having unsafe ops after sum is fine
|
||||
helper_linearizer_opt(a.sum().exp(), [[Opt(OptOps.PADTO, 0, 32)],])
|
||||
helper_linearizer_opt(a.sum(0).exp(), [[Opt(OptOps.PADTO, 1, 32)],])
|
||||
|
||||
def test_padto_sum_not_ok(self):
|
||||
N = 18
|
||||
# NOTE: this setup prevents 17 * 17 contiguous merged into one dimension
|
||||
a = Tensor.rand(N, N).shrink(((0, 17), (0, 17))).exp()
|
||||
# exp is not safe to pad
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(a.exp().sum(), [[Opt(OptOps.PADTO, 0, 32)],])
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(a.exp().sum(0), [[Opt(OptOps.PADTO, 1, 32)],])
|
||||
|
||||
b = a < 1
|
||||
# lt is not safe to pad
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(b.sum(), [[Opt(OptOps.PADTO, 0, 32)],])
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(b.sum(0), [[Opt(OptOps.PADTO, 1, 32)],])
|
||||
|
||||
def test_padto_max(self):
|
||||
N = 18
|
||||
# NOTE: this setup prevents 17 * 17 contiguous merged into one axis
|
||||
a = -Tensor.rand(N, N).shrink(((0, 17), (0, 17))) * 100
|
||||
|
||||
helper_linearizer_opt(a.max(0), [
|
||||
[Opt(OptOps.PADTO, 0, 32)],
|
||||
[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
|
||||
])
|
||||
helper_linearizer_opt(a.max(1), [
|
||||
[Opt(OptOps.PADTO, 0, 32)],
|
||||
[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
|
||||
])
|
||||
|
||||
# cannot pad max kernel on reduce
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(a.max(), [[Opt(OptOps.PADTO, 0, 32)],])
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(a.max(0), [[Opt(OptOps.PADTO, 1, 32)],])
|
||||
|
||||
def test_padto_where(self):
|
||||
Tensor.manual_seed(0)
|
||||
N = 17
|
||||
a = (Tensor.randn(N, N).realize().max(axis=0, keepdim=True) > 1).where(1, 0)
|
||||
helper_linearizer_opt(a.max(0), [
|
||||
[Opt(OptOps.PADTO, 0, 32)],
|
||||
[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
|
||||
])
|
||||
|
||||
def test_padto_where_multioutput(self):
|
||||
Tensor.manual_seed(0)
|
||||
N = 17
|
||||
r = Tensor.randn(N, N).realize().max(axis=0, keepdim=True) > 1
|
||||
a0 = r.where(1, 0)
|
||||
a1 = r.where(2, 0)
|
||||
helper_linearizer_opt([a0.max(0), a1.max(0)], [
|
||||
[Opt(OptOps.PADTO, 0, 32)],
|
||||
[Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
|
||||
])
|
||||
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
|
||||
def test_color_shapes_with_local(self):
|
||||
N = 32
|
||||
Tensor.manual_seed(1552)
|
||||
a = Tensor.rand(N, N)
|
||||
b = Tensor.rand(N, N)
|
||||
r = a@b
|
||||
opts_shapes = [
|
||||
([Opt(OptOps.LOCAL, 0, 2)], [("blue",16),("blue",32),("cyan",2),("red",32)]),
|
||||
([Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.GROUP, 0, 2)], [("blue",16),("blue",32),("cyan",2),("green",2),("red",16)]),
|
||||
# check to ensure local_dims are stable for full UNROLL of the first reduce
|
||||
([Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.UNROLL, 0, 0)], [("blue",16),("blue",32),("cyan",2),("magenta",32)]),
|
||||
([Opt(OptOps.UNROLL, 0, 0),Opt(OptOps.LOCAL, 0, 2)], [("blue",16),("blue",32),("cyan",2),("magenta",32)]),
|
||||
# check behavior for full UNROLL on an existing GROUP
|
||||
([Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.GROUP, 0, 0),Opt(OptOps.UNROLL, 0, 2)], [("blue",16),("blue",32),("cyan",2),("green",16),("magenta",2)]),
|
||||
([Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.GROUP, 0, 0),Opt(OptOps.UNROLL, 0, 0)], [("blue",16),("blue",32),("cyan",2),("magenta",32)]),
|
||||
([Opt(OptOps.GROUP, 0, 0),Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.UNROLL, 0, 0)], [("blue",16),("blue",32),("cyan",2),("magenta",32)]),
|
||||
([Opt(OptOps.GROUP, 0, 2),Opt(OptOps.UNROLL, 0, 0)], [("blue",32),("blue",32),("red",16),("magenta",2)]),
|
||||
]
|
||||
helper_linearizer_opt(r, [x[0] for x in opts_shapes], color_sizes=[x[1] for x in opts_shapes])
|
||||
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "test requires float4")
|
||||
def test_arange_opts(self):
|
||||
a = Tensor.arange(128)
|
||||
# NOTE: arange no longer has reduce ops available for opt
|
||||
helper_linearizer_opt(a, [
|
||||
#[Opt(OptOps.GROUP, 0, 32)],
|
||||
#[Opt(OptOps.GROUPTOP, 0, 32)],
|
||||
[Opt(op=OptOps.LOCAL, axis=0, arg=8)],
|
||||
[Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.UPCAST, axis=0, arg=0)],
|
||||
#[Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.GROUP, axis=0, arg=8)],
|
||||
#[Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.GROUP, axis=0, arg=8), Opt(op=OptOps.UNROLL, axis=1, arg=4)], # noqa: E501
|
||||
])
|
||||
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_threads, "test requires threads")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.global_max is not None and
|
||||
Device[Device.DEFAULT].renderer.global_max[0] > 1, "test requires multicore")
|
||||
def test_thread_opts(self):
|
||||
a = Tensor.rand(4, 4, 4, 4)
|
||||
b = Tensor.rand(4, 4, 4)
|
||||
r = (b.sqrt() + ((a+1).sum(axis=3).exp()))
|
||||
helper_linearizer_opt(r, [
|
||||
[Opt(OptOps.THREAD, 0, 2)],
|
||||
[Opt(OptOps.UPCAST, 0, 2), Opt(OptOps.THREAD, 0, 2)],
|
||||
[Opt(OptOps.UPCAST, 0, 2), Opt(OptOps.THREAD, 0, 2), Opt(OptOps.UNROLL, 0, 2)],
|
||||
] + [[Opt(OptOps.THREAD, 0, 4)] if Device[Device.DEFAULT].renderer.global_max[0] >= 4 else []]
|
||||
+ [[Opt(OptOps.THREAD, 0, 8)] if Device[Device.DEFAULT].renderer.global_max[0] >= 8 else []])
|
||||
|
||||
def test_double_sum_group(self):
|
||||
a = Tensor.rand(4, 4, 4)
|
||||
r = a.sum((1, 2)).sum()
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(r, [[Opt(OptOps.GROUPTOP, 0, 16)],])
|
||||
r = a.sum((1, 2)).sum()
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 1, 4), Opt(OptOps.GROUPTOP, 0, 16)],])
|
||||
r = a.sum((1, 2)).sum()
|
||||
with self.assertRaises(KernelOptError):
|
||||
helper_linearizer_opt(r, [[Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.GROUPTOP, 0, 16)],])
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
208
tinygrad_repo/test/opt/test_tensor_cores.py
Normal file
208
tinygrad_repo/test/opt/test_tensor_cores.py
Normal file
@@ -0,0 +1,208 @@
|
||||
import numpy as np
|
||||
import unittest
|
||||
|
||||
from tinygrad import Device, Tensor, dtypes
|
||||
from tinygrad.tensor import _to_np_dtype
|
||||
from tinygrad.uop.ops import Ops, UOp, buffers
|
||||
from tinygrad.dtype import DType
|
||||
from tinygrad.device import Buffer
|
||||
from tinygrad.helpers import DEV, Context
|
||||
from test.helpers import slow, replace_opts
|
||||
from tinygrad.engine.realize import run_linear
|
||||
from tinygrad.codegen import to_program
|
||||
from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
|
||||
from tinygrad.codegen.opt.tc import amd_cdna_1616128
|
||||
|
||||
# TODO: write a clean version of this
|
||||
from test.backend.test_linearizer import helper_realized_ast, helper_linearizer_opt
|
||||
|
||||
# NOTE: to_program always passes in Device[Device.DEFAULT].renderer explicitly for process_replay!!!
|
||||
|
||||
AMX = "AMX" in DEV.arch
|
||||
|
||||
def run_program(prg:UOp, bufs:list[Buffer]):
|
||||
buf_uops = [UOp.new_buffer(b.device, b.size, b.dtype) for b in bufs]
|
||||
for u,b in zip(buf_uops, bufs): buffers[u] = b
|
||||
run_linear(UOp(Ops.LINEAR, src=(prg.call(*buf_uops),)))
|
||||
|
||||
def helper_tc_ensure_uops_and_opts_count(N: int, M:int, K:int, dtype_in:DType, dtype_out:DType, axis:int=0, tc_select:int=-1, tc_opt:int=0,
|
||||
ensure_triggered:bool=True):
|
||||
a, b = Tensor.rand(M, K, dtype=dtype_in), Tensor.rand(K, N, dtype=dtype_in)
|
||||
r = a.matmul(b, dtype=dtype_out)
|
||||
sched = r.schedule_linear()
|
||||
realized_ast = sched.src[-1].src[0]
|
||||
opts_to_apply = [Opt(OptOps.TC, axis, (tc_select, tc_opt, 1))]
|
||||
|
||||
if ensure_triggered:
|
||||
program = to_program(replace_opts(realized_ast, opts_to_apply), Device[Device.DEFAULT].renderer)
|
||||
wmmas = len([uop for uop in tuple(program.src[2].src) if uop.op is Ops.WMMA])
|
||||
tcs = len([x for x in program.src[0].arg.applied_opts if x.op is OptOps.TC])
|
||||
assert wmmas > 0, "tensor core not triggered"
|
||||
assert tcs == 1, "tensor core opt not included"
|
||||
else:
|
||||
try:
|
||||
program = to_program(replace_opts(realized_ast, opts_to_apply), Device[Device.DEFAULT].renderer)
|
||||
assert False, "OptOps.TC triggered, expected KernelOptError"
|
||||
except KernelOptError: pass
|
||||
|
||||
def helper_tc_allclose(N:int, M:int, K:int, dtype_in:DType, dtype_out:DType, axis:int=0, tc_select:int=-1, tc_opt:int=0, use_tensor_cores:int=1):
|
||||
a, b = Tensor.rand(M, K, dtype=dtype_in), Tensor.rand(K, N, dtype=dtype_in)
|
||||
np_a, np_b = a.numpy(), b.numpy()
|
||||
r = a.matmul(b, dtype=dtype_out)
|
||||
if dtype_in == dtypes.bfloat16: r = r.float()
|
||||
realized_ast, bufs = helper_realized_ast(r)
|
||||
opts = [Opt(op=OptOps.TC, axis=axis, arg=(tc_select, tc_opt, use_tensor_cores))]
|
||||
ast = replace_opts(realized_ast, opts)
|
||||
pu = to_program(ast, Device[Device.DEFAULT].renderer)
|
||||
if use_tensor_cores == 1: assert len([uop for uop in pu.src[2].src if uop.op is Ops.WMMA]) > 0, "wmma not triggered"
|
||||
assert len([x for x in pu.src[0].arg.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
|
||||
run_program(ast, bufs)
|
||||
if dtype_in == dtypes.half: tc_atol, tc_rtol = 1e-2, 1e-3
|
||||
elif dtype_in == dtypes.bfloat16: tc_atol, tc_rtol = (1e-1, 2e-2) if dtype_out == dtypes.bfloat16 else (1e-2, 1e-2)
|
||||
else: tc_atol, tc_rtol = 5e-3, 1e-4
|
||||
c = bufs[0].numpy().reshape((M,N))
|
||||
np.testing.assert_allclose(c, np_a @ np_b, atol=tc_atol, rtol=tc_rtol)
|
||||
|
||||
class TestTensorCores(unittest.TestCase):
|
||||
# TODO: don't skip bf16 for real device (METAL, AMD)
|
||||
@Context(ALLOW_TF32=1)
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
def test_tensor_cores(self):
|
||||
for tc in Device[Device.DEFAULT].renderer.tensor_cores:
|
||||
# for AMX, tc.dims[2] == 1 so reduceop is None thus tensor_cores are not triggered
|
||||
helper_tc_allclose(tc.dims[0], tc.dims[1], 2 if AMX else tc.dims[2], tc.dtype_in, tc.dtype_out, axis=0, tc_opt=0)
|
||||
|
||||
@Context(ALLOW_TF32=1)
|
||||
@unittest.skipIf(Device.DEFAULT == "PYTHON", "not generated on EMULATED device")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
def test_tensor_cores_codegen(self):
|
||||
for tc in Device[Device.DEFAULT].renderer.tensor_cores:
|
||||
n, m, k = tc.dims[0], tc.dims[1], 2 if AMX else tc.dims[2]
|
||||
a, b = Tensor.rand(m, k, dtype=tc.dtype_in), Tensor.rand(k, n, dtype=tc.dtype_in)
|
||||
r = a.matmul(b, dtype=tc.dtype_out)
|
||||
prg = to_program(replace_opts(r.schedule_linear().src[-1].src[0],
|
||||
[Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))]), Device[Device.DEFAULT].renderer)
|
||||
if Device.DEFAULT == "CPU" and DEV.renderer == "LLVM":
|
||||
assert "0x201000" in prg.src[3].arg
|
||||
elif Device.DEFAULT == "AMD" and DEV.renderer == "LLVM":
|
||||
assert "@llvm.amdgcn.wmma" in prg.src[3].arg
|
||||
elif Device[Device.DEFAULT].renderer.suffix == "PTX":
|
||||
assert "mma.sync.aligned" in prg.src[3].arg
|
||||
else:
|
||||
assert "__WMMA_" in prg.src[3].arg
|
||||
|
||||
@Context(ALLOW_TF32=1)
|
||||
@unittest.skipIf((Device.DEFAULT == "AMD") or (Device.DEFAULT == "PYTHON" and Device.default.renderer.target.device == "AMD"), "broken for AMD")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
def test_tensor_cores_padded(self):
|
||||
for tc in Device[Device.DEFAULT].renderer.tensor_cores:
|
||||
helper_tc_allclose(tc.dims[0]+(pad:=1), tc.dims[1]+pad, tc.dims[2]+pad, tc.dtype_in, tc.dtype_out, tc_opt=2)
|
||||
|
||||
# AMD compiler bug: AMD miscompiles non-zero padded tc kernels with -O3, producing wrong results, nans or hang (see #9606)
|
||||
# Internal bug: zero-stride dimensions combined with a mask may produce wrong index/valid for pad == 1 on AMD
|
||||
@unittest.skipUnless((Device.DEFAULT == "AMD") or (Device.DEFAULT == "PYTHON" and Device.default.renderer.target.device == "AMD"),
|
||||
"test for AMD's tc")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
@unittest.skip("warp elements not duplicated properly across lanes")
|
||||
def test_tensor_cores_padded_amd(self):
|
||||
for tc in Device[Device.DEFAULT].renderer.tensor_cores:
|
||||
helper_tc_allclose(tc.dims[0]+(pad:=1), tc.dims[1]+pad, tc.dims[2]+pad, tc.dtype_in, tc.dtype_out, tc_opt=2)
|
||||
|
||||
@Context(ALLOW_TF32=1)
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
def test_tensor_cores_padded_uops(self):
|
||||
for tc in Device[Device.DEFAULT].renderer.tensor_cores:
|
||||
pad = 1
|
||||
|
||||
# check that TC is triggered for TC_OPT=2
|
||||
helper_tc_ensure_uops_and_opts_count(tc.dims[0]+pad, tc.dims[1]+pad, tc.dims[2]+pad,
|
||||
tc.dtype_in, tc.dtype_out, tc_opt=2, ensure_triggered=True)
|
||||
|
||||
# check that TC is not triggered for TC_OPT<2
|
||||
helper_tc_ensure_uops_and_opts_count(tc.dims[0]+pad, tc.dims[1]+pad, tc.dims[2]+pad,
|
||||
tc.dtype_in, tc.dtype_out, tc_opt=1, ensure_triggered=False)
|
||||
helper_tc_ensure_uops_and_opts_count(tc.dims[0]+pad, tc.dims[1]+pad, tc.dims[2]+pad,
|
||||
tc.dtype_in, tc.dtype_out, tc_opt=0, ensure_triggered=False)
|
||||
|
||||
# check excessive padding doesn't trigger padded TC in TC_OPT=2
|
||||
helper_tc_ensure_uops_and_opts_count(tc.dims[0]//4, tc.dims[1], tc.dims[2], tc.dtype_in, tc.dtype_out, tc_opt=2, ensure_triggered=False)
|
||||
helper_tc_ensure_uops_and_opts_count(tc.dims[0], tc.dims[1]//4, tc.dims[2], tc.dtype_in, tc.dtype_out, tc_opt=2, ensure_triggered=False)
|
||||
if not AMX and tc not in amd_cdna_1616128: # AMX tc.dims[2] == 1
|
||||
helper_tc_ensure_uops_and_opts_count(tc.dims[0], tc.dims[1], tc.dims[2]//8, tc.dtype_in, tc.dtype_out, tc_opt=2, ensure_triggered=False)
|
||||
|
||||
@Context(ALLOW_TF32=1)
|
||||
@unittest.skipIf(Device.DEFAULT == "PYTHON", "not generated on EMULATED device")
|
||||
@slow
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
def test_tensor_cores_multi_reduce(self):
|
||||
for tc in Device[Device.DEFAULT].renderer.tensor_cores:
|
||||
if tc.dtype_in is dtypes.bfloat16: continue # <-- broken with numpy
|
||||
# this will be a M=G16, N=G32, M=G16, M=G16, K=R16, K=R16, K=R16 with 9 choices of TC MNK axes
|
||||
golden_result = None
|
||||
for axis in range(9):
|
||||
a = Tensor.rand(16, 16, 29, 29, dtype=tc.dtype_in).realize()
|
||||
b = Tensor.rand(32, 16, 16, 16, dtype=tc.dtype_in).realize()
|
||||
c = a.conv2d(b, padding=1, dtype=tc.dtype_out)
|
||||
realized_ast, real_bufs = helper_realized_ast(c)
|
||||
|
||||
ast = replace_opts(realized_ast, [Opt(OptOps.TC, axis, (-1, 2, 1))])
|
||||
program = to_program(ast, Device[Device.DEFAULT].renderer)
|
||||
assert len([uop for uop in tuple(program.src[2].src) if uop.op is Ops.WMMA]) > 0, "tensor core not triggered"
|
||||
assert len([x for x in program.src[0].arg.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
|
||||
|
||||
# TODO: support this even if numpy doesn't
|
||||
if _to_np_dtype(real_bufs[0].dtype) is None: continue
|
||||
real_bufs[0].copyin(np.zeros((real_bufs[0].size, ), dtype=_to_np_dtype(real_bufs[0].dtype)).data) # Zero to check that all values are filled
|
||||
run_program(ast, real_bufs)
|
||||
result = np.frombuffer(real_bufs[0].as_memoryview(), _to_np_dtype(real_bufs[0].dtype))
|
||||
|
||||
# ensure the results for each choice of axis matches
|
||||
if golden_result is None: golden_result = np.frombuffer(real_bufs[0].as_memoryview(), _to_np_dtype(real_bufs[0].dtype))
|
||||
np.testing.assert_allclose(result, golden_result, atol=0.1, rtol=0.2)
|
||||
|
||||
@Context(ALLOW_TF32=1)
|
||||
@unittest.skipIf(Device.DEFAULT == "PYTHON", "slow on EMULATED device")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
def test_tensor_cores_unroll_phi(self):
|
||||
tc = Device[Device.DEFAULT].renderer.tensor_cores[0]
|
||||
x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
|
||||
r = x.matmul(y, dtype=tc.dtype_out)
|
||||
opts = [Opt(OptOps.UNROLL, 0, 4)]
|
||||
ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
|
||||
for u in tuple(to_program(replace_opts(ast, opts), Device[Device.DEFAULT].renderer).src[2].src):
|
||||
if u.op is Ops.WMMA:
|
||||
assert u.src[-1].src[0].op != Ops.STORE
|
||||
|
||||
@Context(ALLOW_TF32=1)
|
||||
@unittest.skipIf(Device.DEFAULT == "PYTHON", "slow on EMULATED device")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
@unittest.skipIf(Device.DEFAULT in {"CPU"}, "CPU does not support using a different type for accumulation")
|
||||
def test_tensor_cores_unroll_casted_phi(self):
|
||||
tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out][0]
|
||||
x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
|
||||
r = x.matmul(y, dtype=tc.dtype_out)
|
||||
opts = [Opt(OptOps.UNROLL, 0, 4)]
|
||||
ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
|
||||
for u in tuple(to_program(replace_opts(ast, opts), Device[Device.DEFAULT].renderer).src[2].src):
|
||||
if u.op is Ops.WMMA:
|
||||
#assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
|
||||
assert u.src[-1].src[0].op != Ops.STORE
|
||||
|
||||
@Context(ALLOW_TF32=1)
|
||||
@unittest.skipIf(Device.DEFAULT == "PYTHON", "slow on EMULATED device")
|
||||
@unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
|
||||
@unittest.skipIf(Device.DEFAULT in {"CPU"}, "CPU does not support using a different type for accumulation")
|
||||
def test_tensor_cores_unroll_casted_phi_with_children(self):
|
||||
# all STORE children are outside the loop
|
||||
tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out][0]
|
||||
x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
|
||||
r = x.matmul(y, dtype=tc.dtype_out).relu()
|
||||
opts = [Opt(OptOps.UNROLL, 0, 4)]
|
||||
ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
|
||||
for u in tuple(to_program(replace_opts(ast, opts), Device[Device.DEFAULT].renderer).src[2].src):
|
||||
if u.op is Ops.WMMA:
|
||||
#assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
|
||||
assert u.src[-1].src[0].op != Ops.STORE
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user