Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[bf16] add bf16 kernel: sigmoid & sqrt & softplus & square #40004

Merged
merged 8 commits into from
Mar 7, 2022
Merged

[bf16] add bf16 kernel: sigmoid & sqrt & softplus & square #40004

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
d3089a7
add activ
zhangbo9674 Feb 28, 2022
6659459
refine unittest
zhangbo9674 Mar 1, 2022
e97b752
refine unittest
zhangbo9674 Mar 1, 2022
80b1d27
refine unittest
zhangbo9674 Mar 2, 2022
215fc9d
Merge branch 'develop' of /~https://github.com/PaddlePaddle/Paddle into…
zhangbo9674 Mar 2, 2022
9340324
refine unittest
zhangbo9674 Mar 2, 2022
61707fa
Merge branch 'develop' of /~https://github.com/PaddlePaddle/Paddle into…
zhangbo9674 Mar 4, 2022
e7f8a07
refine code
zhangbo9674 Mar 4, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 24 additions & 7 deletions paddle/fluid/operators/activation_op.kps
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1509,15 +1509,19 @@ namespace plat = paddle::platform;
ops::ActivationCudaKernel<paddle::platform::CUDADeviceContext, \
ops::functor<double>>, \
ops::ActivationCudaKernel<plat::CUDADeviceContext, \
ops::functor<plat::float16>>); \
ops::functor<plat::float16>>, \
ops::ActivationCudaKernel<plat::CUDADeviceContext, \
ops::functor<plat::bfloat16>>); \
REGISTER_OP_CUDA_KERNEL( \
act_type##_grad, \
ops::ActivationGradCudaKernel<plat::CUDADeviceContext, \
ops::grad_functor<float>>, \
ops::ActivationGradCudaKernel<plat::CUDADeviceContext, \
ops::grad_functor<double>>, \
ops::ActivationGradCudaKernel<plat::CUDADeviceContext, \
ops::grad_functor<plat::float16>>);
ops::grad_functor<plat::float16>>, \
ops::ActivationGradCudaKernel<plat::CUDADeviceContext, \
ops::grad_functor<plat::bfloat16>>);

#define REGISTER_ACTIVATION_CUDA_KERNEL_INT(act_type, op_name, functor, \
grad_functor) \
Expand All @@ -1531,7 +1535,9 @@ namespace plat = paddle::platform;
ops::ActivationCudaKernel<paddle::platform::CUDADeviceContext, \
ops::functor<int64_t>>, \
ops::ActivationCudaKernel<plat::CUDADeviceContext, \
ops::functor<plat::float16>>); \
ops::functor<plat::float16>>, \
ops::ActivationCudaKernel<plat::CUDADeviceContext, \
ops::functor<plat::bfloat16>>); \
REGISTER_OP_CUDA_KERNEL( \
act_type##_grad, \
ops::ActivationGradCudaKernel<plat::CUDADeviceContext, \
Expand All @@ -1543,7 +1549,9 @@ namespace plat = paddle::platform;
ops::ActivationGradCudaKernel<plat::CUDADeviceContext, \
ops::grad_functor<int64_t>>, \
ops::ActivationGradCudaKernel<plat::CUDADeviceContext, \
ops::grad_functor<plat::float16>>);
ops::grad_functor<plat::float16>>, \
ops::ActivationGradCudaKernel<plat::CUDADeviceContext, \
ops::grad_functor<plat::bfloat16>>);

/* ======================== leaky relu register ============================ */
REGISTER_ACTIVATION_CUDA_KERNEL(leaky_relu, LeakyRelu, CudaLeakyReluFunctor,
Expand Down Expand Up @@ -1650,7 +1658,9 @@ REGISTER_OP_CUDA_KERNEL(
ops::SigmoidDoubleGradKernel<paddle::platform::CUDADeviceContext,
ops::SigmoidGradGradFunctor<double>>,
ops::SigmoidDoubleGradKernel<plat::CUDADeviceContext,
ops::SigmoidGradGradFunctor<plat::float16>>);
ops::SigmoidGradGradFunctor<plat::float16>>,
ops::SigmoidDoubleGradKernel<plat::CUDADeviceContext,
ops::SigmoidGradGradFunctor<plat::bfloat16>>);

REGISTER_OP_CUDA_KERNEL(
sigmoid_triple_grad,
Expand All @@ -1659,7 +1669,10 @@ REGISTER_OP_CUDA_KERNEL(
ops::SigmoidTripleGradKernel<paddle::platform::CUDADeviceContext,
ops::SigmoidTripleGradFunctor<double>>,
ops::SigmoidTripleGradKernel<plat::CUDADeviceContext,
ops::SigmoidTripleGradFunctor<plat::float16>>);
ops::SigmoidTripleGradFunctor<plat::float16>>,
ops::SigmoidTripleGradKernel<
plat::CUDADeviceContext,
ops::SigmoidTripleGradFunctor<plat::bfloat16>>);
/* ========================================================================== */

/* =========================== tanh register ============================ */
Expand Down Expand Up @@ -1696,7 +1709,9 @@ REGISTER_OP_CUDA_KERNEL(
ops::SqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
ops::SqrtGradGradFunctor<double>>,
ops::SqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
ops::SqrtGradGradFunctor<plat::float16>>);
ops::SqrtGradGradFunctor<plat::float16>>,
ops::SqrtDoubleGradKernel<paddle::platform::CUDADeviceContext,
ops::SqrtGradGradFunctor<plat::bfloat16>>);
/* ========================================================================== */

/* =========================== rsqrt register =============================
Expand Down Expand Up @@ -1726,6 +1741,8 @@ REGISTER_OP_CUDA_KERNEL(
ops::SquareGradGradFunctor<double>>,
ops::SquareDoubleGradKernel<plat::CUDADeviceContext,
ops::SquareGradGradFunctor<plat::float16>>,
ops::SquareDoubleGradKernel<plat::CUDADeviceContext,
ops::SquareGradGradFunctor<plat::bfloat16>>,
ops::SquareDoubleGradKernel<paddle::platform::CUDADeviceContext,
ops::SquareGradGradFunctor<int>>,
ops::SquareDoubleGradKernel<paddle::platform::CUDADeviceContext,
Expand Down
7 changes: 7 additions & 0 deletions paddle/fluid/operators/amp/fp16_type_traits.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@ limitations under the License. */

#pragma once

#include "paddle/fluid/platform/bfloat16.h"
#include "paddle/fluid/platform/float16.h"

namespace paddle {
Expand All @@ -32,6 +33,12 @@ class MPTypeTrait<platform::float16> {
using Type = float;
};

template <>
class MPTypeTrait<platform::bfloat16> {
public:
using Type = float;
};

} // namespace details
} // namespace operators
} // namespace paddle
3 changes: 2 additions & 1 deletion paddle/fluid/operators/dropout_impl.cu.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -266,7 +266,8 @@ void DropoutFwGPUKernelDriver(const platform::CUDADeviceContext& dev_ctx,
cudaMemcpyDeviceToDevice, stream));
#endif
} else {
T factor = static_cast<T>(1.0f - dropout_prob);
using MT = typename details::MPTypeTrait<T>::Type;
MT factor = static_cast<MT>(1.0f - dropout_prob);
std::vector<const framework::Tensor*> ins = {&x};
std::vector<framework::Tensor*> outs = {y};
auto functor = phi::funcs::ScaleFunctor<T>(factor);
Expand Down
4 changes: 4 additions & 0 deletions paddle/phi/common/bfloat16.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -310,6 +310,10 @@ HOSTDEVICE inline bool(isfinite)(const bfloat16& a) {
return !((isnan)(a)) && !((isinf)(a));
}

HOSTDEVICE inline bfloat16(abs)(const bfloat16& a) {
return bfloat16(std::abs(static_cast<float>(a)));
}

inline std::ostream& operator<<(std::ostream& os, const bfloat16& a) {
os << static_cast<float>(a);
return os;
Expand Down
113 changes: 113 additions & 0 deletions python/paddle/fluid/tests/unittests/test_activation_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,6 +183,34 @@ def test_check_grad(self):
self.check_grad(['X'], 'Out', max_relative_error=0.01)


@unittest.skipIf(not core.is_compiled_with_cuda(),
"core is not compiled with CUDA")
class TestSigmoidBF16(OpTest):
def setUp(self):
self.op_type = "sigmoid"
self.init_dtype()

np.random.seed(1024)
x = np.random.uniform(-1, 1, [11, 17]).astype(np.float32)
out = 1 / (1 + np.exp(-x))

self.inputs = {
'X': OpTest.np_dtype_to_fluid_dtype(convert_float_to_uint16(x))
}
self.outputs = {'Out': convert_float_to_uint16(out)}

def init_dtype(self):
self.dtype = np.uint16

def test_check_output(self):
place = core.CUDAPlace(0)
self.check_output_with_place(place)

def test_check_grad(self):
place = core.CUDAPlace(0)
self.check_grad_with_place(place, ['X'], 'Out')


class TestSilu(TestActivation):
def setUp(self):
self.op_type = "silu"
Expand Down Expand Up @@ -945,6 +973,34 @@ def test_check_grad(self):
self.check_grad(['X'], 'Out')


@unittest.skipIf(not core.is_compiled_with_cuda(),
"core is not compiled with CUDA")
class TestSqrtBF16(OpTest):
def setUp(self):
self.op_type = "sqrt"
self.init_dtype()

np.random.seed(1023)
x = np.random.uniform(0.1, 1, [11, 17]).astype(np.float32)
out = np.sqrt(x)

self.inputs = {
'X': OpTest.np_dtype_to_fluid_dtype(convert_float_to_uint16(x))
}
self.outputs = {'Out': convert_float_to_uint16(out)}

def init_dtype(self):
self.dtype = np.uint16

def test_check_output(self):
place = core.CUDAPlace(0)
self.check_output_with_place(place)

def test_check_grad(self):
place = core.CUDAPlace(0)
self.check_grad_with_place(place, ['X'], 'Out')


class TestRsqrt(TestActivation):
def setUp(self):
self.op_type = "rsqrt"
Expand Down Expand Up @@ -2195,6 +2251,34 @@ def test_check_grad(self):
self.check_grad(['X'], 'Out', max_relative_error=0.007)


@unittest.skipIf(not core.is_compiled_with_cuda(),
"core is not compiled with CUDA")
class TestSquareBF16(OpTest):
def setUp(self):
self.op_type = "square"
self.init_dtype()

np.random.seed(1024)
x = np.random.uniform(0.1, 1, [11, 17]).astype(np.float32)
out = np.square(x)

self.inputs = {
'X': OpTest.np_dtype_to_fluid_dtype(convert_float_to_uint16(x))
}
self.outputs = {'Out': convert_float_to_uint16(out)}

def init_dtype(self):
self.dtype = np.uint16

def test_check_output(self):
place = core.CUDAPlace(0)
self.check_output_with_place(place)

def test_check_grad(self):
place = core.CUDAPlace(0)
self.check_grad_with_place(place, ['X'], 'Out', numeric_grad_delta=0.5)


class TestPow(TestActivation):
def setUp(self):
self.op_type = "pow"
Expand Down Expand Up @@ -2433,6 +2517,35 @@ def test_check_grad(self):
self.check_grad(['X'], 'Out')


@unittest.skipIf(not core.is_compiled_with_cuda(),
"core is not compiled with CUDA")
class TestSoftplusBF16(OpTest):
def setUp(self):
self.op_type = "softplus"
self.init_dtype()

beta = 2
threshold = 15

np.random.seed(1024)
x = np.random.uniform(-1, 1, [10, 12]).astype(np.float32)
out = ref_softplus(x, beta, threshold)
self.inputs = {'X': convert_float_to_uint16(x)}
self.attrs = {'beta': beta, "threshold": threshold}
self.outputs = {'Out': convert_float_to_uint16(out)}

def init_dtype(self):
self.dtype = np.uint16

def test_check_output(self):
place = core.CUDAPlace(0)
self.check_output_with_place(place)

def test_check_grad(self):
place = core.CUDAPlace(0)
self.check_grad_with_place(place, ['X'], 'Out', numeric_grad_delta=0.05)


class TestSoftplusAPI(unittest.TestCase):
# test paddle.nn.Softplus, paddle.nn.functional.softplus
def setUp(self):
Expand Down