Nigel
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

gpu Gelu kernel support fp16

This commit is contained in:
wilfChen 2020-06-18 19:33:27 +08:00
parent 971f10d222
commit 8f4cd76582
5 changed files with 139 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

97
mindspore/ccsrc/kernel/gpu/cuda_impl/gelu_impl.cu
View File

@ -14,32 +14,62 @@
* limitations under the License.
*/
#include "kernel/gpu/cuda_impl/gelu_impl.cuh"
#include "device/gpu/cuda_common.h"
template<typename T>
__global__ void GeluKernel(size_t size, T* input_addr, T* output_addr) {
template <typename T>
__global__ void GeluKernel(size_t size, T *input_addr, T *output_addr) {
// formula:
// gelu(x) = 0.5 * x * (1.0 + tanh(y))
// tanh(y) = 2 / (1 + exp(-2y)) - 1)
// y = sqrt(2/pi) * (x + 0.044715 * x^3)
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < (size); pos += blockDim.x * gridDim.x) {
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < size; pos += blockDim.x * gridDim.x) {
float x = input_addr[pos];
float tanh_res = tanh(0.7978845608 * (x + 0.044715 * x * x * x));
output_addr[pos] = 0.5 * x * (1.0 + tanh_res);
}
}
template<typename T>
void Gelu(size_t size, T* input_addr, T* output_addr, cudaStream_t cuda_stream) {
template <>
__global__ void GeluKernel(size_t size, half *input_addr, half *output_addr) {
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < size; pos += blockDim.x * gridDim.x) {
half x = input_addr[pos];
float tanh_res = tanh(__half2float(half(0.7978845608) * (x + half(0.044715) * x * x * x)));
output_addr[pos] = half(0.5) * x * (half(1.0) + __float2half(tanh_res));
}
}
template <>
__global__ void GeluKernel(size_t size, half2 *input_addr, half2 *output_addr) {
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < size; pos += blockDim.x * gridDim.x) {
half2 x = input_addr[pos];
float2 tanh_param = __half22float2(half2(0.7978845608, 0.7978845608) * (x + half2(0.044715, 0.044715) * x * x * x));
float2 tanh_res;
tanh_res.x = tanh(tanh_param.x);
tanh_res.y = tanh(tanh_param.y);
output_addr[pos] = half2(0.5, 0.5) * x * (half2(1.0, 1.0) + __float22half2_rn(tanh_res));
}
}
template <typename T>
void Gelu(size_t size, T *input_addr, T *output_addr, cudaStream_t cuda_stream) {
GeluKernel<<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(size, input_addr, output_addr);
return;
}
template <>
void Gelu(size_t size, half *input_addr, half *output_addr, cudaStream_t cuda_stream) {
if (size % 2 == 0) {
GeluKernel<half2><<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(
size / 2, reinterpret_cast<half2 *>(input_addr), reinterpret_cast<half2 *>(output_addr));
} else {
GeluKernel<half><<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(size, input_addr, output_addr);
}
return;
}
template<typename T>
__global__ void GeluGradKernel(size_t size, T* dy_addr, T* x_addr, T* dx_addr) {
template <typename T>
__global__ void GeluGradKernel(size_t size, T *dy_addr, T *x_addr, T *dx_addr) {
// formula:
// dx = dy * y'
// y' = 0.5 * (1 + tanh(tanh_para)) +
@ -50,16 +80,57 @@ __global__ void GeluGradKernel(size_t size, T* dy_addr, T* x_addr, T* dx_addr) {
T x = x_addr[pos];
T tanh_res = tanh(0.7978845608 * (x + 0.044715 * x * x * x));
T mul_right = 0.7978845608 + 0.1070322244 * x * x;
T y_res = 0.5 * (1 + tanh_res) + 0.5 * x * (1 - tanh_res * tanh_res) * mul_right;
T y_res = 0.5 * (1.0 + tanh_res) + 0.5 * x * (1.0 - tanh_res * tanh_res) * mul_right;
dx_addr[pos] = dy_addr[pos] * y_res;
}
}
template<typename T>
void GeluGradKernel(size_t size, T* dy_addr, T* x_addr, T* dx_addr, cudaStream_t cuda_stream) {
template <typename T>
__global__ void GeluGradKernel(size_t size, half2 *dy_addr, half2 *x_addr, half2 *dx_addr) {
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < (size); pos += blockDim.x * gridDim.x) {
half2 x = x_addr[pos];
float2 tanh_param = __half22float2(half2(0.7978845608, 0.7978845608) * (x + half2(0.044715, 0.044715) * x * x * x));
float2 tanh_res;
tanh_res.x = tanh(tanh_param.x);
tanh_res.y = tanh(tanh_param.y);
half2 tanh_res_half = __float22half2_rn(tanh_res);
half2 mul_right = half2(0.7978845608, 0.7978845608) + half2(0.1070322244, 0.1070322244) * x * x;
half2 y_res = half2(0.5, 0.5) * (half2(1.0, 1.0) + tanh_res_half) +
half2(0.5, 0.5) * x * (half2(1.0, 1.0) - tanh_res_half * tanh_res_half) * mul_right;
dx_addr[pos] = dy_addr[pos] * y_res;
}
}
template <typename T>
__global__ void GeluGradKernel(size_t size, half *dy_addr, half *x_addr, half *dx_addr) {
for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < (size); pos += blockDim.x * gridDim.x) {
half x = x_addr[pos];
half tanh_param = half(0.7978845608) * (x + half(0.044715) * x * x * x);
half tanh_res = __float2half_rn(tanh(__half2float(tanh_param)));
half mul_right = half(0.7978845608) + half(0.1070322244) * x * x;
half y_res = half(0.5) * (half(1.0) + tanh_res) + half(0.5) * x * (half(1.0) - tanh_res * tanh_res) * mul_right;
dx_addr[pos] = dy_addr[pos] * y_res;
}
}
template <typename T>
void GeluGradKernel(size_t size, T *dy_addr, T *x_addr, T *dx_addr, cudaStream_t cuda_stream) {
GeluGradKernel<<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(size, dy_addr, x_addr, dx_addr);
}
template <>
void GeluGradKernel(size_t size, half *dy_addr, half *x_addr, half *dx_addr, cudaStream_t cuda_stream) {
if (size % 2 == 0) {
GeluGradKernel<half2><<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(
size / 2, reinterpret_cast<half2 *>(dy_addr), reinterpret_cast<half2 *>(x_addr),
reinterpret_cast<half2 *>(dx_addr));
} else {
GeluGradKernel<half><<<GET_BLOCKS(size), GET_THREADS, 0, cuda_stream>>>(size, dy_addr, x_addr, dx_addr);
}
return;
}
template void Gelu(size_t size, float* input_addr, float* output_addr, cudaStream_t cuda_stream);
template void GeluGradKernel(size_t size, float* dy_addr, float* x_addr, float* dx_addr, cudaStream_t cuda_stream);
template void Gelu(size_t size, float *input_addr, float *output_addr, cudaStream_t cuda_stream);
template void Gelu(size_t size, half *input_addr, half *output_addr, cudaStream_t cuda_stream);
template void GeluGradKernel(size_t size, float *dy_addr, float *x_addr, float *dx_addr, cudaStream_t cuda_stream);
template void GeluGradKernel(size_t size, half *dy_addr, half *x_addr, half *dx_addr, cudaStream_t cuda_stream);

7
mindspore/ccsrc/kernel/gpu/nn/gelu_grad_kernel.cc
View File

@ -25,5 +25,12 @@ MS_REG_GPU_KERNEL_ONE(GeluGrad,
.AddInputAttr(kNumberTypeFloat32)
.AddOutputAttr(kNumberTypeFloat32),
GeLUGpuGradKernel, float)
MS_REG_GPU_KERNEL_ONE(GeluGrad,
KernelAttr()
.AddInputAttr(kNumberTypeFloat16)
.AddInputAttr(kNumberTypeFloat16)
.AddInputAttr(kNumberTypeFloat16)
.AddOutputAttr(kNumberTypeFloat16),
GeLUGpuGradKernel, half)
} // namespace kernel
} // namespace mindspore

2
mindspore/ccsrc/kernel/gpu/nn/gelu_kernel.cc
View File

@ -20,5 +20,7 @@ namespace mindspore {
namespace kernel {
MS_REG_GPU_KERNEL_ONE(Gelu, KernelAttr().AddInputAttr(kNumberTypeFloat32).AddOutputAttr(kNumberTypeFloat32),
GeluGpuKernel, float)
MS_REG_GPU_KERNEL_ONE(Gelu, KernelAttr().AddInputAttr(kNumberTypeFloat16).AddOutputAttr(kNumberTypeFloat16),
GeluGpuKernel, half)
} // namespace kernel
} // namespace mindspore

32
tests/st/ops/gpu/test_gelu_grad_op.py
View File

@ -58,7 +58,37 @@ def test_gelugrad():
grad = Grad(net)
output = grad(x_ms, dy_ms)
print(output)
expect = [0.50963277, 0.9414753, 0.2667653, 0.21358444, 0.25243032, 0.0352667,
0.34266686, 0.57757664, 0.04707306, 0.51536125]
assert np.allclose(output[0].asnumpy(), expect)
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_gelugrad_fp16():
np.random.seed(42)
x_np = np.random.randn(5, 3, 6).astype(np.float16)
dy_np = np.random.randn(5, 3, 6).astype(np.float16)
net = GeluNet()
grad = Grad(net)
output = grad(Tensor(x_np), Tensor(dy_np))
expect = [[[8.4045e-02, 3.7817e-01, -6.6748e-01, -3.6914e-01, -1.2415e-01, -4.6362e-01],
[3.3301e-01, 2.6270e-01, 7.7534e-04, -2.0947e-01, -2.2021e-01, -6.4880e-02],
[-2.3633e-01, 7.6538e-02, 1.8280e-02, 3.8635e-02, -1.6235e-01, 1.2964e-01]],
[[-1.4801e-02, 9.6130e-03, -2.1660e+00, -8.5602e-03, 3.3356e-02, -3.1885e-01],
[-2.0355e-02, 1.7737e-01, 3.8719e-03, -9.1895e-01, 8.4717e-02, 2.0593e-01],
[5.8350e-02, -1.0020e+00, 6.8652e-01, 1.3428e-01, 6.0352e-01, -2.6270e-01]],
[[-6.5820e-01, 5.1147e-02, -1.2650e-02, -3.2983e-01, -1.5410e+00, 4.3518e-02],
[-4.3359e-01, 1.2659e-01, 1.1792e-01, 2.2705e-02, -1.2329e-01, -3.5278e-01],
[6.2109e-01, 1.3611e-01, 1.7041e-01, 2.7124e-01, -5.5908e-02, 1.7212e-01]],
[[2.8320e-01, 8.3252e-01, 4.2480e-02, -3.4473e-01, 3.9429e-01, 3.1958e-01],
[3.6499e-02, 1.2250e-01, 7.1350e-02, -2.7267e-02, 3.0029e-01, -8.0566e-01],
[8.2617e-01, 5.1367e-01, -9.2480e-01, 3.3203e-02, -7.5684e-01, 8.8623e-01]],
[[5.4590e-01, -9.2383e-01, -2.8107e-02, 4.2432e-01, 4.6826e-01, 5.0879e-01],
[-1.4062e-01, 6.6284e-02, -2.9126e-01, -6.3086e-01, -8.6975e-02, 4.1504e-02],
[-6.3171e-03, 1.0852e-01, 1.3779e-02, 1.0947e+00, -3.0334e-02, 2.3828e+00]]]
assert np.allclose(output[0].asnumpy(), expect, rtol=1e-2)

13
tests/st/ops/gpu/test_gelu_op.py
View File

@ -91,3 +91,16 @@ def test_gelu_neg():
y_ms = net(x_ms)
assert np.allclose(y_np, y_ms.asnumpy())
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_gelu_4d_fp16():
x_np = np.random.random((32, 3, 224, 224)).astype(np.float16)
y_np = GeluCompute(x_np)
x_ms = Tensor(x_np)
net = GeluNet()
y_ms = net(x_ms)
assert np.allclose(y_np, y_ms.asnumpy(), rtol=1e-3)