def test_reduce_map(target, dev, ref_data, in_shape, axis, keepdims,
reduce_type, dtype):
target = tvm.target.Target(target)
if target.kind.name == "vulkan" and reduce_type in ["sum", "any", "all"]:
pytest.xfail(f"Vulkan backend has known errors on {reduce_type}")
in_npy, in_npy_map, out_npy = ref_data
# Build the logic and compile the function
A = te.placeholder(shape=in_shape, name="A", dtype=dtype)
A1 = topi.sqrt(topi.exp(A))
out_dtype = dtype
if reduce_type == "sum":
B = topi.sum(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "all":
B = topi.all(A, axis=axis, keepdims=keepdims)
elif reduce_type == "any":
B = topi.any(A, axis=axis, keepdims=keepdims)
elif reduce_type == "max":
B = topi.max(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "min":
B = topi.min(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "argmax":
B = topi.argmax(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
elif reduce_type == "argmin":
B = topi.argmin(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
else:
raise NotImplementedError
with tvm.target.Target(target):
s = tvm.topi.testing.get_reduce_schedule(target)(B)
foo = tvm.build(s, [A, B], target, name=reduce_type)
data_tvm = tvm.nd.array(in_npy, device=dev)
out_tvm = tvm.nd.empty(shape=out_npy.shape, device=dev, dtype=out_dtype)
foo(data_tvm, out_tvm)
if reduce_type == "argmax" or reduce_type == "argmin":
out_tvm_indices = out_tvm.numpy()
if keepdims:
out_tvm_indices = np.take(out_tvm_indices, indices=0, axis=axis)
if axis is None:
out_tvm_val = in_npy_map.ravel()[out_tvm_indices]
else:
other_indices = tuple(
np.indices(in_shape[0:axis] + in_shape[(axis + 1):]))
sel_indices = other_indices[0:axis] + (
out_tvm_indices, ) + other_indices[axis:]
out_tvm_val = in_npy_map[sel_indices]
if reduce_type == "argmax":
tvm.testing.assert_allclose(out_tvm_val, in_npy_map.max(axis=axis),
1e-3, 1e-3)
elif reduce_type == "argmin":
tvm.testing.assert_allclose(out_tvm_val, in_npy_map.min(axis=axis),
1e-3, 1e-3)
else:
tvm.testing.assert_allclose(out_tvm.numpy(), out_npy, 1e-3, 1e-3)
def batch_norm_bwd(N, C, H, W, dtype="float32"):
dshape = (N, C, H, W)
oshape = (C, )
bshape = (1, C, 1, 1)
sshape = (1, )
data = te.placeholder(dshape, name="data", dtype=dtype)
scale = te.placeholder(oshape, name="scale", dtype=dtype)
saved_mean = te.placeholder(oshape, name="saved_mean", dtype=dtype)
saved_var = te.placeholder(oshape, name="saved_var", dtype=dtype)
eps = te.placeholder(sshape, name="eps", dtype=dtype)
grad_output = te.placeholder(dshape, name="data", dtype=dtype)
axis = (0, 2, 3)
num_ele = dshape[0] * dshape[2] * dshape[3]
frac_num_ele = 1.0 / num_ele
# compute grad_input
mean_sum = topi.sum(data, axis, True)
mean = topi.multiply(mean_sum, frac_num_ele)
var_sub = topi.subtract(data, mean)
var_mul = topi.multiply(var_sub, var_sub)
var_sum = topi.sum(var_mul, axis, True)
var = topi.multiply(var_sum, frac_num_ele)
var_eps = topi.add(var, eps)
output_sqrt = topi.sqrt(var_eps)
x_norm = topi.subtract(data, mean)
x_hat = topi.divide(x_norm, output_sqrt)
dx_hat = topi.multiply(grad_output, topi.reshape(scale, bshape))
grad_input_sum1 = topi.sum(dx_hat * x_hat, axis, True)
grad_input_sum2 = topi.sum(dx_hat, axis, True)
grad_input_left = topi.divide(frac_num_ele, topi.sqrt(var_eps))
grad_input_right1 = topi.subtract(topi.multiply(dx_hat, num_ele),
grad_input_sum2)
grad_input_right2 = topi.multiply(x_hat, grad_input_sum1)
grad_input = topi.multiply(
grad_input_left, topi.subtract(grad_input_right1, grad_input_right2))
# compute grad_scale and grad_bias
grad_scale = topi.sum(grad_output * x_hat, axis)
grad_bias = topi.sum(grad_output, axis)
return [
data, scale, saved_mean, saved_var, eps, grad_output, grad_input,
grad_scale, grad_bias
]
def batch_norm_fwd(N, C, H, W, dtype="float32"):
dshape = (N, C, H, W)
oshape = (C, )
bshape = (1, C, 1, 1)
sshape = (1, )
data = te.placeholder(dshape, name="data", dtype=dtype)
scale = te.placeholder(oshape, name="scale", dtype=dtype)
bias = te.placeholder(oshape, name="bias", dtype=dtype)
running_mean = te.placeholder(oshape, name="running_mean", dtype=dtype)
running_var = te.placeholder(oshape, name="running_var", dtype=dtype)
eps = te.placeholder(sshape, name="eps", dtype=dtype)
momentum = te.placeholder(sshape, name="momentum", dtype=dtype)
axis = (0, 2, 3)
num_ele = dshape[0] * dshape[2] * dshape[3]
frac_num_ele = 1.0 / num_ele
# compute batch mean
mean_sum = topi.sum(data, axis, keepdims=True)
saved_mean = topi.multiply(mean_sum, frac_num_ele)
# compute batch rvars
var_sub = topi.subtract(data, saved_mean)
var_mul = topi.multiply(var_sub, var_sub)
var_sum = topi.sum(var_mul, axis, keepdims=True)
var = topi.multiply(var_sum, frac_num_ele)
output_add = topi.add(var, eps)
saved_rvars = topi.sqrt(output_add)
# # compute output
output_sub = topi.subtract(data, saved_mean)
output_norm = topi.divide(output_sub, saved_rvars)
scale_board = topi.reshape(scale, bshape)
bias_board = topi.reshape(bias, bshape)
output = topi.add(topi.multiply(output_norm, scale_board), bias_board)
# reshape saved_rvars
saved_rvars = topi.reshape(saved_rvars, oshape)
# update running mean
running_mean_mul1 = topi.multiply(running_mean,
topi.subtract(1.0, momentum))
running_mean_mul2 = topi.multiply(topi.reshape(saved_mean, oshape),
momentum)
running_mean_out = topi.add(running_mean_mul1, running_mean_mul2)
# update running var
saved_var_mul1 = topi.multiply(running_var, topi.subtract(1.0, momentum))
saved_var_mul2 = topi.multiply(topi.reshape(var, oshape), momentum)
running_var_out = topi.add(saved_var_mul1, saved_var_mul2)
# reshape saved_mean
saved_mean = topi.reshape(saved_mean, oshape)
return [
data, scale, bias, running_mean, running_var, momentum, eps, output,
saved_mean, saved_rvars, running_mean_out, running_var_out
]
def verify_reduce_map_ele(in_shape,
axis,
keepdims,
type="sum",
dtype="float32"):
# Build the logic and compile the function
A = te.placeholder(shape=in_shape, name="A", dtype=dtype)
A1 = topi.sqrt(topi.exp(A))
out_dtype = dtype
if type == "sum":
B = topi.sum(A1, axis=axis, keepdims=keepdims)
elif type == "all":
B = topi.all(A, axis=axis, keepdims=keepdims)
elif type == "any":
B = topi.any(A, axis=axis, keepdims=keepdims)
elif type == "max":
B = topi.max(A1, axis=axis, keepdims=keepdims)
elif type == "min":
B = topi.min(A1, axis=axis, keepdims=keepdims)
elif type == "argmax":
B = topi.argmax(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
elif type == "argmin":
B = topi.argmin(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
else:
raise NotImplementedError
def check_device(device, ctx):
print("Running on target: %s" % device)
with tvm.target.Target(device):
s = tvm.topi.testing.get_reduce_schedule(device)(B)
foo = tvm.build(s, [A, B], device, name=type)
# Test
if dtype == "bool":
in_npy_map = in_npy = np.random.choice([True, False],
size=in_shape)
else:
in_npy = np.random.uniform(-1, 1, size=in_shape).astype(dtype)
in_npy_map = np.sqrt(np.exp(in_npy)).astype(dtype)
if type == "sum":
out_npy = in_npy_map.sum(axis=axis, keepdims=keepdims)
elif type == "all" and dtype == "bool":
out_npy = in_npy_map.all(axis=axis, keepdims=keepdims)
elif type == "any" and dtype == "bool":
out_npy = in_npy_map.any(axis=axis, keepdims=keepdims)
elif type == "max":
out_npy = in_npy_map.max(axis=axis, keepdims=keepdims)
elif type == "min":
out_npy = in_npy_map.min(axis=axis, keepdims=keepdims)
elif type == "argmax":
out_npy = _my_npy_argmax(in_npy_map, axis=axis, keepdims=keepdims)
elif type == "argmin":
out_npy = _my_npy_argmin(in_npy_map, axis=axis, keepdims=keepdims)
else:
raise NotImplementedError
data_tvm = tvm.nd.array(in_npy, ctx=ctx)
out_tvm = tvm.nd.empty(shape=out_npy.shape, ctx=ctx, dtype=out_dtype)
for _ in range(1):
foo(data_tvm, out_tvm)
if type == "argmax" or type == "argmin":
out_tvm_indices = out_tvm.asnumpy()
if keepdims:
out_tvm_indices = np.take(out_tvm_indices,
indices=0,
axis=axis)
if axis is None:
out_tvm_val = in_npy_map.ravel()[out_tvm_indices]
else:
other_indices = tuple(
np.indices(in_shape[0:axis] + in_shape[(axis + 1):]))
sel_indices = other_indices[0:axis] + (
out_tvm_indices, ) + other_indices[axis:]
out_tvm_val = in_npy_map[sel_indices]
if type == "argmax":
tvm.testing.assert_allclose(out_tvm_val,
in_npy_map.max(axis=axis), 1e-3,
1e-3)
elif type == "argmin":
tvm.testing.assert_allclose(out_tvm_val,
in_npy_map.min(axis=axis), 1e-3,
1e-3)
else:
tvm.testing.assert_allclose(out_tvm.asnumpy(), out_npy, 1e-3, 1e-3)
for device, ctx in tvm.testing.enabled_targets():
check_device(device, ctx)
def test_reduce_map(hexagon_session: Session, ref_data, in_shape, axis,
keepdims, reduce_type, dtype):
in_npy, in_npy_map, out_npy = ref_data
# Build the logic and compile the function
A = te.placeholder(shape=in_shape, name="A", dtype=dtype)
A1 = topi.sqrt(topi.exp(A))
out_dtype = dtype
if reduce_type == "sum":
B = topi.sum(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "all":
B = topi.all(A, axis=axis, keepdims=keepdims)
elif reduce_type == "any":
B = topi.any(A, axis=axis, keepdims=keepdims)
elif reduce_type == "max":
B = topi.max(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "min":
B = topi.min(A1, axis=axis, keepdims=keepdims)
elif reduce_type == "argmax":
B = topi.argmax(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
elif reduce_type == "argmin":
B = topi.argmin(A1, axis=axis, keepdims=keepdims)
out_dtype = "int32"
else:
raise NotImplementedError
target_hexagon = tvm.target.hexagon("v68")
with tvm.target.Target(target_hexagon):
fschedule = topi.hexagon.schedule_reduce
s = fschedule(B)
func = tvm.build(s, [A, B],
tvm.target.Target(target_hexagon, host=target_hexagon),
name=reduce_type)
mod = hexagon_session.load_module(func)
dev = hexagon_session.device
data_tvm = tvm.nd.array(in_npy, device=dev)
out_tvm = tvm.nd.empty(shape=out_npy.shape, device=dev, dtype=out_dtype)
mod[reduce_type](data_tvm, out_tvm)
if reduce_type == "argmax" or reduce_type == "argmin":
out_tvm_indices = out_tvm.numpy()
if keepdims:
out_tvm_indices = np.take(out_tvm_indices, indices=0, axis=axis)
if axis is None:
out_tvm_val = in_npy_map.ravel()[out_tvm_indices]
else:
other_indices = tuple(
np.indices(in_shape[0:axis] + in_shape[(axis + 1):]))
sel_indices = other_indices[0:axis] + (
out_tvm_indices, ) + other_indices[axis:]
out_tvm_val = in_npy_map[sel_indices]
if reduce_type == "argmax":
tvm.testing.assert_allclose(out_tvm_val, in_npy_map.max(axis=axis),
1e-3, 1e-3)
elif reduce_type == "argmin":
tvm.testing.assert_allclose(out_tvm_val, in_npy_map.min(axis=axis),
1e-3, 1e-3)
else:
tvm.testing.assert_allclose(out_tvm.numpy(), out_npy, 1e-3, 1e-3)