forked from mindspore-Ecosystem/mindspore
217 lines
6.5 KiB
Python
217 lines
6.5 KiB
Python
# Copyright 2022 Huawei Technologies Co., Ltd
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# ============================================================================
|
|
|
|
import time
|
|
import numpy as np
|
|
import pytest
|
|
|
|
import mindspore.context as context
|
|
import mindspore.nn as nn
|
|
from mindspore import Tensor
|
|
from mindspore.ops import composite as C
|
|
from mindspore.ops import operations as P
|
|
from mindspore.ops import functional as F
|
|
from mindspore.ops.functional import vmap
|
|
from mindspore.common.api import ms_function
|
|
|
|
context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
|
|
|
|
|
|
def fast_gelu_grad_compute(x, dy):
|
|
"""FastGeluGradCompute."""
|
|
div_up = np.exp(-1.702 * x) + 1.702 * x * np.exp(-1.702 * x) + np.exp(1.702 * (x - np.abs(x)))
|
|
div_down = (np.exp(-1.702 * x) + 1) ** 2
|
|
return dy * div_up / div_down
|
|
|
|
|
|
def fast_gelu_compute(x):
|
|
"""FastGeluCompute."""
|
|
return x * np.exp(0.851 * (x - np.abs(x))) / (1 + np.exp(-1.702 * np.abs(x)))
|
|
|
|
|
|
class FastGeluNet(nn.Cell):
|
|
"""FastGeluNet."""
|
|
|
|
def __init__(self):
|
|
"""Init."""
|
|
super(FastGeluNet, self).__init__()
|
|
self.fast_gelu = P.FastGeLU()
|
|
|
|
def construct(self, x):
|
|
"""Construct."""
|
|
return self.fast_gelu(x)
|
|
|
|
|
|
class FastGeLUGrad(nn.Cell):
|
|
"""FastGeLUGrad."""
|
|
|
|
def __init__(self, network):
|
|
"""Init."""
|
|
super(FastGeLUGrad, self).__init__()
|
|
self.fast_gelu_grad = C.GradOperation(get_all=True, sens_param=True)
|
|
self.network = network
|
|
|
|
def construct(self, input_data, sens):
|
|
"""Construct."""
|
|
gout = self.fast_gelu_grad(self.network)(input_data, sens)
|
|
return gout
|
|
|
|
|
|
def np_all_close_with_loss(out, expect):
|
|
"""np_all_close_with_loss"""
|
|
return np.allclose(out, expect, 0.005, 0.005, equal_nan=True)
|
|
|
|
|
|
@pytest.mark.level0
|
|
@pytest.mark.platform_x86_gpu_training
|
|
@pytest.mark.env_onecard
|
|
@pytest.mark.parametrize('shape', [(2,), (4, 5), (3, 4, 5, 6)])
|
|
@pytest.mark.parametrize('dtype', [np.float32, np.float16])
|
|
def test_fast_gelu_grad(shape, dtype):
|
|
"""
|
|
Feature: FastGeLUGrad gpu kernel
|
|
Description: test the rightness of FastGeLUGrad gpu kernel.
|
|
Expectation: Success.
|
|
"""
|
|
prop = 1 if np.random.random() > 0.5 else -1
|
|
dy_np = (np.random.randn(*shape) * prop).astype(dtype)
|
|
x_np = (np.random.randn(*shape) * prop).astype(dtype)
|
|
expect = fast_gelu_grad_compute(dy_np, x_np)
|
|
|
|
dy_ms = Tensor(dy_np)
|
|
x_ms = Tensor(x_np)
|
|
net = FastGeluNet()
|
|
grad = FastGeLUGrad(net)
|
|
output = grad(dy_ms, x_ms)
|
|
assert np_all_close_with_loss(output[0].asnumpy(), expect)
|
|
|
|
|
|
@pytest.mark.level0
|
|
@pytest.mark.platform_x86_gpu_training
|
|
@pytest.mark.env_onecard
|
|
@pytest.mark.parametrize('shape', [(2,), (4, 5), (3, 4, 5, 6)])
|
|
@pytest.mark.parametrize('dtype', [np.float32, np.float16])
|
|
def test_fast_gelu(shape, dtype):
|
|
"""
|
|
Feature: FastGeLU gpu kernel
|
|
Description: test the rightness of FastGeLU gpu kernel.
|
|
Expectation: Success.
|
|
"""
|
|
prop = 100 if np.random.random() > 0.5 else -100
|
|
x_np = (np.random.randn(*shape) * prop).astype(dtype)
|
|
y_np = fast_gelu_compute(x_np)
|
|
|
|
x_ms = Tensor(x_np)
|
|
net = FastGeluNet()
|
|
y_ms = net(x_ms)
|
|
assert np_all_close_with_loss(y_np, y_ms.asnumpy())
|
|
|
|
x_ms = Tensor(x_np)
|
|
y_fun = F.fast_gelu(x_ms)
|
|
assert np_all_close_with_loss(y_np, y_fun.asnumpy())
|
|
|
|
x_ms = Tensor(x_np)
|
|
fast_gelu_nn = nn.FastGelu()
|
|
y_nn = fast_gelu_nn(x_ms)
|
|
assert np_all_close_with_loss(y_np, y_nn.asnumpy())
|
|
|
|
|
|
@pytest.mark.level0
|
|
@pytest.mark.platform_x86_gpu_training
|
|
@pytest.mark.env_onecard
|
|
@pytest.mark.parametrize('dtype', [np.float32, np.float16])
|
|
def test_fast_gelu_grad_vmap(dtype, shape=(100, 2)):
|
|
"""
|
|
Feature: FastGeLUGrad gpu kernel
|
|
Description: test the rightness of FastGeLUGrad gpu kernel vmap feature.
|
|
Expectation: Success.
|
|
"""
|
|
net = FastGeluNet()
|
|
grad = FastGeLUGrad(net)
|
|
|
|
def fast_gelu_grad_func(dy, x):
|
|
"""fast_gelu_grad_func"""
|
|
output = grad(dy, x)
|
|
return output[0]
|
|
|
|
prop = 1 if np.random.random() > 0.5 else -1
|
|
dy_np = (np.random.randn(*shape) * prop).astype(dtype)
|
|
x_np = (np.random.randn(*shape) * prop).astype(dtype)
|
|
dy = Tensor(dy_np)
|
|
x = Tensor(x_np)
|
|
dy = F.sub(dy, 0)
|
|
x = F.sub(x, 0)
|
|
|
|
start_time = time.perf_counter()
|
|
output_vmap = vmap(fast_gelu_grad_func, in_axes=(0, 0))(dy, x)
|
|
vmap_time = time.perf_counter() - start_time
|
|
|
|
start_time_manually = time.perf_counter()
|
|
|
|
@ms_function
|
|
def manually_batched(dys, xs):
|
|
"""manually_batched"""
|
|
output = []
|
|
for i in range(dys.shape[0]):
|
|
output.append(fast_gelu_grad_func(dys[i], xs[i]))
|
|
return F.stack(output)
|
|
|
|
output_manually = manually_batched(dy, x)
|
|
manually_time = time.perf_counter() - start_time_manually
|
|
|
|
assert np_all_close_with_loss(output_vmap.asnumpy(), output_manually.asnumpy())
|
|
assert vmap_time < manually_time
|
|
|
|
|
|
@pytest.mark.level0
|
|
@pytest.mark.platform_x86_gpu_training
|
|
@pytest.mark.env_onecard
|
|
@pytest.mark.parametrize('dtype', [np.float32, np.float16])
|
|
def test_fast_gelu_vmap(dtype, shape=(100, 2)):
|
|
"""
|
|
Feature: FastGeLU gpu kernel
|
|
Description: test the rightness of FastGeLU gpu kernel vmap feature.
|
|
Expectation: Success.
|
|
"""
|
|
|
|
def fast_gelu_func(x):
|
|
"""fast_gelu_func"""
|
|
return P.FastGeLU()(x)
|
|
|
|
prop = 100 if np.random.random() > 0.5 else -100
|
|
x_np = (np.random.randn(*shape) * prop).astype(dtype)
|
|
x = Tensor(x_np)
|
|
x = F.sub(x, 0)
|
|
|
|
start_time = time.perf_counter()
|
|
output_vmap = vmap(fast_gelu_func, in_axes=(0,))(x)
|
|
vmap_time = time.perf_counter() - start_time
|
|
|
|
start_time_manually = time.perf_counter()
|
|
|
|
@ms_function
|
|
def manually_batched(xs):
|
|
"""manually_batched"""
|
|
output = []
|
|
for i in range(xs.shape[0]):
|
|
output.append(fast_gelu_func(xs[i]))
|
|
return F.stack(output)
|
|
|
|
output_manually = manually_batched(x)
|
|
manually_time = time.perf_counter() - start_time_manually
|
|
|
|
assert np_all_close_with_loss(output_vmap.asnumpy(), output_manually.asnumpy())
|
|
assert vmap_time < manually_time
|