Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 87667c66 编写于 作者: W wbn 提交者： GitHub
浏览文件
操作

Add the new XDNN implementation. test=kunlun (#42683) 

* Add the new XDNN implementation. test=kunlun

* Add the new XDNN implementation. test=kunlun

* Modify the code based on review, test=kunlun
上级 34cda80b
隐藏空白更改
内联 并排
Showing with 179 addition and 261 deletion
cmake/external/xpu.cmake
浏览文件 @ 87667c66
......@@ -9,7 +9,7 @@ SET(XPU_RT_LIB_NAME "libxpurt.so")
if(NOT DEFINED XPU_BASE_URL)
SET(XPU_BASE_URL_WITHOUT_DATE "https://baidu-kunlun-product.cdn.bcebos.com/KL-SDK/klsdk-dev")
SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20220510")
SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20220511")
else()
SET(XPU_BASE_URL "${XPU_BASE_URL}")
endif()
......@@ -17,7 +17,7 @@ endif()
# ubuntu and centos: use output by XDNN API team
if(NOT DEFINED XPU_XDNN_BASE_URL)
SET(XPU_XDNN_BASE_URL_WITHOUT_DATE "https://klx-sdk-release-public.su.bcebos.com/xdnn/dev")
SET(XPU_XDNN_BASE_URL "${XPU_XDNN_BASE_URL_WITHOUT_DATE}/20220510")
SET(XPU_XDNN_BASE_URL "${XPU_XDNN_BASE_URL_WITHOUT_DATE}/20220511")
else()
SET(XPU_XDNN_BASE_URL "${XPU_XDNN_BASE_URL}")
endif()
......
paddle/fluid/operators/log_loss_op_xpu.cc
浏览文件 @ 87667c66
......@@ -12,6 +12,7 @@ limitations under the License. */
#include <memory>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/device/device_wrapper.h"
namespace paddle {
namespace operators {
......@@ -21,67 +22,47 @@ template <typename DeviceContext, typename T, typename AttrType = T>
class LogLossXPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
/*** TODO wait XDNN new interface
auto* predict = ctx.Input<Tensor>("Predicted");
auto* labels = ctx.Input<Tensor>("Labels");
auto* loss = ctx.Output<Tensor>("Loss");
auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
loss->mutable_data<T>(ctx.GetPlace());
int n = predict->numel();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
int r =
xpu::log_loss_fwd(dev_ctx.x_context(), n, epsilon,
predict->data<T>(),
labels->data<T>(), loss->data<T>());
PADDLE_ENFORCE_EQ(
r, xpu::Error_t::SUCCESS,
platform::errors::External(
"XPU log_loss kernel return wrong value[%d], please check
whether "
"Baidu Kunlun Card is properly installed.",
r));
***/
auto* predict = ctx.Input<Tensor>("Predicted");
auto* labels = ctx.Input<Tensor>("Labels");
auto* loss = ctx.Output<Tensor>("Loss");
auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
loss->mutable_data<T>(ctx.GetPlace());
int n = predict->numel();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
int r = xpu::log_loss(dev_ctx.x_context(), predict->data<T>(),
labels->data<T>(), loss->data<T>(), n, epsilon);
PADDLE_ENFORCE_XDNN_SUCCESS(r, "log_loss");
}
};
template <typename DeviceContext, typename T, typename AttrType = T>
class LogLossGradXPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
/*** TODO wait XDNN new interface
auto* predict = ctx.Input<Tensor>("Predicted");
auto* labels = ctx.Input<Tensor>("Labels");
auto* dloss = ctx.Input<Tensor>(framework::GradVarName("Loss"));
auto* dpred = ctx.Output<Tensor>(framework::GradVarName("Predicted"));
if (!dpred) {
return;
}
auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
dpred->mutable_data<T>(ctx.GetPlace());
int n = predict->numel();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
int r = xpu::log_loss_bwd(dev_ctx.x_context(), n, epsilon,
predict->data<T>(), labels->data<T>(),
dloss->data<T>(), dpred->data<T>());
PADDLE_ENFORCE_EQ(
r, xpu::Error_t::SUCCESS,
platform::errors::External(
"XPU log_loss kernel return wrong value[%d], please check
whether "
"Baidu Kunlun Card is properly installed.",
r));
***/
auto* predict = ctx.Input<Tensor>("Predicted");
auto* labels = ctx.Input<Tensor>("Labels");
auto* dloss = ctx.Input<Tensor>(framework::GradVarName("Loss"));
auto* dpred = ctx.Output<Tensor>(framework::GradVarName("Predicted"));
if (!dpred) {
return;
}
auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
dpred->mutable_data<T>(ctx.GetPlace());
int n = predict->numel();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
int r = xpu::log_loss_grad(dev_ctx.x_context(), predict->data<T>(),
labels->data<T>(), dloss->data<T>(),
dpred->data<T>(), n, epsilon);
PADDLE_ENFORCE_XDNN_SUCCESS(r, "log_loss_grad");
}
};
} // namespace operators
} // namespace paddle
// namespace ops = paddle::operators;
// REGISTER_OP_XPU_KERNEL(
// log_loss, ops::LogLossXPUKernel<paddle::platform::XPUDeviceContext,
// float>);
// REGISTER_OP_XPU_KERNEL(
// log_loss_grad,
// ops::LogLossGradXPUKernel<paddle::platform::XPUDeviceContext, float>);
namespace ops = paddle::operators;
REGISTER_OP_XPU_KERNEL(
log_loss, ops::LogLossXPUKernel<paddle::platform::XPUDeviceContext, float>);
REGISTER_OP_XPU_KERNEL(
log_loss_grad,
ops::LogLossGradXPUKernel<paddle::platform::XPUDeviceContext, float>);
#endif
paddle/fluid/operators/optimizers/lamb_op_xpu.cc
浏览文件 @ 87667c66
......@@ -14,6 +14,7 @@ limitations under the License. */
#include "paddle/fluid/operators/optimizers/lamb_op.h"
#include "gflags/gflags.h"
#include "paddle/fluid/platform/device/device_wrapper.h"
namespace paddle {
namespace operators {
......@@ -25,111 +26,75 @@ template <typename DeviceContext, typename T>
class LambOpXPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
/*** TODO wait XDNN new interface
using paddle::framework::LoDTensor;
const auto* param_var = ctx.InputVar("Param");
PADDLE_ENFORCE_EQ(param_var->IsType<framework::LoDTensor>(), true,
platform::errors::InvalidArgument(
"The Var(%s)'s type should be LoDTensor, "
"but the received is %s",
ctx.InputNames("Param").front(),
framework::ToTypeName(param_var->Type())));
using paddle::framework::LoDTensor;
const auto* param_var = ctx.InputVar("Param");
PADDLE_ENFORCE_EQ(param_var->IsType<framework::LoDTensor>(), true,
platform::errors::InvalidArgument(
"The Var(%s)'s type should be LoDTensor, "
"but the received is %s",
ctx.InputNames("Param").front(),
framework::ToTypeName(param_var->Type())));
using paddle::framework::LoDTensor;
using paddle::framework::LoDTensor;
// inputs
T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
T weight_decay = static_cast<T>(ctx.Attr<float>("weight_decay"));
T beta1 = static_cast<T>(ctx.Attr<float>("beta1"));
T beta2 = static_cast<T>(ctx.Attr<float>("beta2"));
auto& param = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Param"), "Input",
"Param", "Lamb");
auto* grad_var = ctx.InputVar("Grad");
auto& mom1 = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Moment1"), "Input",
"Moment1", "Lamb");
auto& mom2 = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Moment2"), "Input",
"Moment2", "Lamb");
auto& lr = GET_DATA_SAFELY(ctx.Input<LoDTensor>("LearningRate"),
"Input",
"LearningRate", "Lamb");
// inputs
T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
T weight_decay = static_cast<T>(ctx.Attr<float>("weight_decay"));
T beta1 = static_cast<T>(ctx.Attr<float>("beta1"));
T beta2 = static_cast<T>(ctx.Attr<float>("beta2"));
auto& param = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Param"), "Input",
"Param", "Lamb");
auto* grad_var = ctx.InputVar("Grad");
auto& mom1 = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Moment1"), "Input",
"Moment1", "Lamb");
auto& mom2 = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Moment2"), "Input",
"Moment2", "Lamb");
auto& lr = GET_DATA_SAFELY(ctx.Input<LoDTensor>("LearningRate"), "Input",
"LearningRate", "Lamb");
auto& beta1_pow = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Beta1Pow"),
"Input",
"Beta1Pow", "Lamb");
auto& beta2_pow = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Beta2Pow"),
"Input",
"Beta2Pow", "Lamb");
auto& beta1_pow = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Beta1Pow"), "Input",
"Beta1Pow", "Lamb");
auto& beta2_pow = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Beta2Pow"), "Input",
"Beta2Pow", "Lamb");
auto& param_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("ParamOut"),
"Output", "ParamOut", "Lamb");
auto& mom1_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("Moment1Out"),
"Output", "Moment1Out", "Lamb");
auto& mom2_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("Moment2Out"),
"Output", "Moment2Out", "Lamb");
auto& beta1_pow_out =
GET_DATA_SAFELY(ctx.Output<LoDTensor>("Beta1PowOut"),
"Output", "Beta1PowOut", "Lamb");
auto& beta2_pow_out =
GET_DATA_SAFELY(ctx.Output<LoDTensor>("Beta2PowOut"),
"Output", "Beta2PowOut", "Lamb");
auto& dev_ctx = ctx.template device_context<DeviceContext>();
auto& param_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("ParamOut"),
"Output", "ParamOut", "Lamb");
auto& mom1_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("Moment1Out"),
"Output", "Moment1Out", "Lamb");
auto& mom2_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("Moment2Out"),
"Output", "Moment2Out", "Lamb");
auto& beta1_pow_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("Beta1PowOut"),
"Output", "Beta1PowOut", "Lamb");
auto& beta2_pow_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("Beta2PowOut"),
"Output", "Beta2PowOut", "Lamb");
auto& dev_ctx = ctx.template device_context<DeviceContext>();
if (grad_var->IsType<framework::LoDTensor>()) {
auto& grad = *ctx.Input<LoDTensor>("Grad");
int r = xpu::lamb(dev_ctx.x_context(), grad.template data<T>(),
mom1.template data<T>(), mom2.template data<T>(),
param.template data<T>(), beta1_pow.template
data<T>(),
beta2_pow.template data<T>(), beta1, beta2, epsilon,
weight_decay, lr.template data<T>(),
mom1_out.template mutable_data<T>(ctx.GetPlace()),
mom2_out.template mutable_data<T>(ctx.GetPlace()),
param_out.template mutable_data<T>(ctx.GetPlace()),
beta1_pow_out.template
mutable_data<T>(ctx.GetPlace()),
beta2_pow_out.template
mutable_data<T>(ctx.GetPlace()),
param.numel());
if (grad_var->IsType<framework::LoDTensor>()) {
auto& grad = *ctx.Input<LoDTensor>("Grad");
int r = xpu::lamb(
dev_ctx.x_context(), grad.template data<T>(), mom1.template data<T>(),
mom2.template data<T>(), param.template data<T>(),
beta1_pow.template data<T>(), beta2_pow.template data<T>(),
mom1_out.template mutable_data<T>(ctx.GetPlace()),
mom2_out.template mutable_data<T>(ctx.GetPlace()),
param_out.template mutable_data<T>(ctx.GetPlace()),
beta1_pow_out.template mutable_data<T>(ctx.GetPlace()),
beta2_pow_out.template mutable_data<T>(ctx.GetPlace()), beta1, beta2,
epsilon, weight_decay, lr.template data<T>(), param.numel());
if (r == xpu::Error_t::INVALID_PARAM) {
PADDLE_ENFORCE_EQ(
r, xpu::Error_t::SUCCESS,
platform::errors::InvalidArgument(
"XPU kernel error of LambOp, error message: INVALID_PARAM, "
"please check your input & output."));
} else if (r == xpu::Error_t::RUNTIME_ERROR) {
PADDLE_ENFORCE_EQ(r, xpu::Error_t::SUCCESS,
platform::errors::Unavailable(
"XPU kernel error of LambOp, error message: "
"RUNTIME_ERROR, please check whether Baidu "
"Kunlun Card is properly installed."));
} else if (r == xpu::Error_t::NO_ENOUGH_WORKSPACE) {
PADDLE_ENFORCE_EQ(r, xpu::Error_t::SUCCESS,
platform::errors::ResourceExhausted(
"XPU kernel error of LambOp, error "
"message: NO_ENOUGH_WORKSPACE, XPU "
"has no enough memory."));
} else {
PADDLE_ENFORCE_EQ(r, xpu::Error_t::SUCCESS,
platform::errors::ResourceExhausted(
"XPU kernel error of LambOp, error "
"message: OTHER "
"XPU API returns error code: %d.",
r));
}
} else {
PADDLE_THROW(platform::errors::InvalidArgument(
"Variable type not supported by lamb_op. Expect LoDTensor, "
"but got %s",
framework::ToTypeName(param_var->Type())));
}
**/
PADDLE_ENFORCE_XDNN_SUCCESS(r, "lamb");
} else {
PADDLE_THROW(platform::errors::InvalidArgument(
"Variable type not supported by lamb_op. Expect LoDTensor, "
"but got %s",
framework::ToTypeName(param_var->Type())));
}
}
};
} // namespace operators
} // namespace paddle
// namespace ops = paddle::operators;
// REGISTER_OP_XPU_KERNEL(
// lamb, ops::LambOpXPUKernel<paddle::platform::XPUDeviceContext, float>);
namespace ops = paddle::operators;
REGISTER_OP_XPU_KERNEL(
lamb, ops::LambOpXPUKernel<paddle::platform::XPUDeviceContext, float>);
#endif
paddle/fluid/operators/optimizers/rmsprop_op_xpu.cc
浏览文件 @ 87667c66
......@@ -17,6 +17,7 @@ limitations under the License. */
#include <gflags/gflags.h>
#include <iostream>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/device/device_wrapper.h"
namespace paddle {
namespace operators {
......@@ -40,122 +41,88 @@ template <typename DeviceContext, typename T>
class RmspropOpXPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
/*** TODO wait XDNN new interface
using paddle::framework::LoDTensor;
// check Param & Grad tensor type
const auto* param_var = ctx.InputVar("Param");
PADDLE_ENFORCE_EQ(param_var->IsType<LoDTensor>(), true,
platform::errors::InvalidArgument(
"Tensor holds the wrong type，Expected Var(%s)'s "
"type is LoDTensor, "
"but the received is %s",
ctx.InputNames("Param").front(),
framework::ToTypeName(param_var->Type())));
const auto* grad_var = ctx.InputVar("Grad");
PADDLE_ENFORCE_EQ(grad_var->IsType<LoDTensor>(), true,
platform::errors::InvalidArgument(
"Tensor holds the wrong type，Expected Var(%s)'s "
"type is LoDTensor, "
"but the received is %s",
ctx.InputNames("Grad").front(),
framework::ToTypeName(grad_var->Type())));
// inputs
auto& param = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Param"), "Input",
"Param", "Rmsprop");
auto& meanSquare = GET_DATA_SAFELY(ctx.Input<LoDTensor>("MeanSquare"),
"Input", "MeanSquare", "Rmsprop");
auto& grad = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Grad"), "Input",
"Grad",
"Rmsprop");
auto& mom = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Moment"), "Input",
"Moment", "Rmsprop");
auto* learning_rate = ctx.Input<Tensor>("LearningRate");
PADDLE_ENFORCE_EQ(learning_rate->dims().size(), 1,
platform::errors::InvalidArgument(
"learining rate should have dimension = 1."
" But received learning rate dim [%s] ",
learning_rate->dims().size()));
T lr = static_cast<T>(GetAttrFromTensor(learning_rate));
// constants
T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
T decay = static_cast<T>(ctx.Attr<float>("decay"));
T momentum = static_cast<T>(ctx.Attr<float>("momentum"));
// outputs
auto& param_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("ParamOut"),
"Output", "ParamOut", "Rmsprop");
auto& mom_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("MomentOut"),
"Output", "MomentOut", "Rmsprop");
auto& mom_sqrt_out =
GET_DATA_SAFELY(ctx.Output<LoDTensor>("MeanSquareOut"),
"Output", "MeanSquareOut",
"Rmsprop");
auto& dev_ctx = ctx.template device_context<DeviceContext>();
///// rmsprop优化算法
///
/// ms_out[i] = rho * ms[i] + (1 - rho) * (g[i] * g[i]);
///
/// mom_out[i] = momentum * mom[i] + lr *
/// (g[i] / ((float)sqrt(ms_out[i] + epsilon)));
///
/// p_out[i] = p[i] - mom_out[i];
/// DLL_EXPORT int rmsprop(Context* ctx, const float* p,
/// const float* ms, const float* g, const float* mom,
/// float epsilon, float rho, float momentum, float lr,
/// float *ms_out, float *mom_out, float *p_out, int n)
int r = xpu::rmsprop(dev_ctx.x_context(), param.template data<T>(),
meanSquare.template data<T>(), grad.template
data<T>(),
mom.template data<T>(), epsilon, decay, momentum,
lr,
mom_sqrt_out.template
mutable_data<T>(ctx.GetPlace()),
mom_out.template mutable_data<T>(ctx.GetPlace()),
param_out.template mutable_data<T>(ctx.GetPlace()),
param.numel());
if (r == xpu::Error_t::INVALID_PARAM) {
PADDLE_ENFORCE_EQ(
r, xpu::Error_t::SUCCESS,
platform::errors::InvalidArgument(
"XPU kernel error of RmspropOp, error message: INVALID_PARAM,
"
"please check your input & output."));
} else if (r == xpu::Error_t::RUNTIME_ERROR) {
PADDLE_ENFORCE_EQ(r, xpu::Error_t::SUCCESS,
platform::errors::Unavailable(
"XPU kernel error of RmspropOp, error message: "
"RUNTIME_ERROR, please check whether Baidu "
"Kunlun Card is properly installed."));
} else if (r == xpu::Error_t::NO_ENOUGH_WORKSPACE) {
PADDLE_ENFORCE_EQ(r, xpu::Error_t::SUCCESS,
platform::errors::ResourceExhausted(
"XPU kernel error of RmspropOp, error "
"message: NO_ENOUGH_WORKSPACE, XPU "
"has no enough memory."));
} else {
PADDLE_ENFORCE_EQ(r, xpu::Error_t::SUCCESS,
platform::errors::ResourceExhausted(
"XPU kernel error of RmspropOp, error "
"message: OTHER "
"XPU API returns error code: %d.",
r));
}
***/
using paddle::framework::LoDTensor;
// check Param & Grad tensor type
const auto* param_var = ctx.InputVar("Param");
PADDLE_ENFORCE_EQ(param_var->IsType<LoDTensor>(), true,
platform::errors::InvalidArgument(
"Tensor holds the wrong type，Expected Var(%s)'s "
"type is LoDTensor, "
"but the received is %s",
ctx.InputNames("Param").front(),
framework::ToTypeName(param_var->Type())));
const auto* grad_var = ctx.InputVar("Grad");
PADDLE_ENFORCE_EQ(grad_var->IsType<LoDTensor>(), true,
platform::errors::InvalidArgument(
"Tensor holds the wrong type，Expected Var(%s)'s "
"type is LoDTensor, "
"but the received is %s",
ctx.InputNames("Grad").front(),
framework::ToTypeName(grad_var->Type())));
// inputs
auto& param = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Param"), "Input",
"Param", "Rmsprop");
auto& meanSquare = GET_DATA_SAFELY(ctx.Input<LoDTensor>("MeanSquare"),
"Input", "MeanSquare", "Rmsprop");
auto& grad = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Grad"), "Input", "Grad",
"Rmsprop");
auto& mom = GET_DATA_SAFELY(ctx.Input<LoDTensor>("Moment"), "Input",
"Moment", "Rmsprop");
auto* learning_rate = ctx.Input<Tensor>("LearningRate");
PADDLE_ENFORCE_EQ(learning_rate->dims().size(), 1,
platform::errors::InvalidArgument(
"learining rate should have dimension = 1."
" But received learning rate dim [%s] ",
learning_rate->dims().size()));
T lr = static_cast<T>(GetAttrFromTensor(learning_rate));
// constants
T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
T decay = static_cast<T>(ctx.Attr<float>("decay"));
T momentum = static_cast<T>(ctx.Attr<float>("momentum"));
// outputs
auto& param_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("ParamOut"),
"Output", "ParamOut", "Rmsprop");
auto& mom_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("MomentOut"),
"Output", "MomentOut", "Rmsprop");
auto& mom_sqrt_out = GET_DATA_SAFELY(ctx.Output<LoDTensor>("MeanSquareOut"),
"Output", "MeanSquareOut", "Rmsprop");
auto& dev_ctx = ctx.template device_context<DeviceContext>();
///// rmsprop优化算法
///
/// ms_out[i] = rho * ms[i] + (1 - rho) * (g[i] * g[i]);
///
/// mom_out[i] = momentum * mom[i] + lr *
/// (g[i] / ((float)sqrt(ms_out[i] + epsilon)));
///
/// p_out[i] = p[i] - mom_out[i];
/// DLL_EXPORT int rmsprop(Context* ctx, const float* p,
/// const float* ms, const float* g, const float* mom,
/// float epsilon, float rho, float momentum, float lr,
/// float *ms_out, float *mom_out, float *p_out, int n)
int r = xpu::rmsprop(dev_ctx.x_context(), grad.template data<T>(),
param.template data<T>(),
meanSquare.template data<T>(), mom.template data<T>(),
param_out.template mutable_data<T>(ctx.GetPlace()),
mom_sqrt_out.template mutable_data<T>(ctx.GetPlace()),
mom_out.template mutable_data<T>(ctx.GetPlace()),
epsilon, decay, momentum, lr, param.numel());
PADDLE_ENFORCE_XDNN_SUCCESS(r, "rmsprop");
}
};
} // namespace operators
} // namespace paddle
// namespace ops = paddle::operators;
// REGISTER_OP_XPU_KERNEL(
// rmsprop,
// ops::RmspropOpXPUKernel<paddle::platform::XPUDeviceContext, float>);
namespace ops = paddle::operators;
REGISTER_OP_XPU_KERNEL(
rmsprop,
ops::RmspropOpXPUKernel<paddle::platform::XPUDeviceContext, float>);
#endif
paddle/fluid/platform/device/xpu/xpu1_op_list.h
浏览文件 @ 87667c66
......@@ -145,6 +145,7 @@ XPUOpMap& get_kl1_ops() {
{"hard_switch", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"iou_similarity",
XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"lamb", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"layer_norm_grad",
XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"layer_norm", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
......@@ -174,6 +175,9 @@ XPUOpMap& get_kl1_ops() {
pOpKernelType(vartype::INT32, XPUPlace()),
pOpKernelType(vartype::INT64, XPUPlace()),
pOpKernelType(vartype::FP32, XPUPlace())})},
{"log_loss_grad",
XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"log_loss", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"logsumexp", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"log", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"lookup_table_v2_grad",
......@@ -232,6 +236,7 @@ XPUOpMap& get_kl1_ops() {
pOpKernelType(vartype::INT32, XPUPlace()),
pOpKernelType(vartype::BOOL, XPUPlace()),
pOpKernelType(vartype::FP32, XPUPlace())})},
{"rmsprop", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"rnn_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"rnn", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
{"roi_align_grad",
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或