Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite

体验新版 GitCode，发现更多精彩内容 >>

未验证 提交 5dac279a 编写于 作者: Z zhupengyang 提交者： GitHub
浏览文件
操作

[XPU] move elementwise and pool uts (#3060)

上级 a34f06a1
隐藏空白更改
内联 并排
Showing with 3 addition and 457 deletion
lite/kernels/xpu/bridges/elementwise_ops_test.cc 已删除 100644 → 0
浏览文件 @ a34f06a1
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// 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.
#include "lite/operators/elementwise_ops.h"
#include <gtest/gtest.h>
#include <random>
#include "lite/core/op_registry.h"
#include "lite/kernels/xpu/bridges/registry.h"
#include "lite/kernels/xpu/bridges/test_helper.h"
namespace paddle {
namespace lite {
namespace kernels {
namespace xpu {
namespace bridges {
template <typename dtype>
void elementwise_add_ref(const std::shared_ptr<operators::ElementwiseOp> op) {
Scope* scope = op->scope();
const OpInfo* op_info = op->op_info();
auto x = scope->FindVar(op_info->Input("X").front())->GetMutable<Tensor>();
auto y = scope->FindVar(op_info->Input("Y").front())->GetMutable<Tensor>();
auto out =
scope->FindVar(op_info->Output("Out").front())->GetMutable<Tensor>();
auto x_data = x->data<dtype>();
auto y_data = y->data<dtype>();
dtype* out_data = out->mutable_data<dtype>();
auto x_dims = x->dims();
auto y_dims = y->dims();
int axis = op_info->GetAttr<int>("axis");
if (axis < 0) {
axis = x_dims.size() - y_dims.size();
}
int batch = 1;
int channels = 1;
int num = 1;
for (int i = 0; i < axis; ++i) {
batch *= x_dims[i];
}
for (int i = 0; i < y_dims.size(); ++i) {
channels *= y_dims[i];
}
for (int i = y_dims.size() + axis; i < x_dims.size(); ++i) {
num *= x_dims[i];
}
// do elementwise add/sub/max...
std::string elt_type = "add";
if (elt_type == "add") {
for (int i = 0; i < batch; ++i) {
for (int j = 0; j < channels; ++j) {
int offset = (i * channels + j) * num;
const dtype* din_ptr = x_data + offset;
const dtype diny_data = y_data[j];
dtype* dout_ptr = out_data + offset;
for (int k = 0; k < num; ++k) {
*dout_ptr = *din_ptr + diny_data;
dout_ptr++;
din_ptr++;
}
}
}
} else if (elt_type == "sub") {
for (int i = 0; i < batch; ++i) {
for (int j = 0; j < channels; ++j) {
int offset = (i * channels + j) * num;
const dtype* din_ptr = x_data + offset;
const dtype diny_data = y_data[j];
dtype* dout_ptr = out_data + offset;
for (int k = 0; k < num; ++k) {
*dout_ptr = *din_ptr - diny_data;
dout_ptr++;
din_ptr++;
}
}
}
} else if (elt_type == "mul") {
for (int i = 0; i < batch; ++i) {
for (int j = 0; j < channels; ++j) {
int offset = (i * channels + j) * num;
const dtype* din_ptr = x_data + offset;
const dtype diny_data = y_data[j];
dtype* dout_ptr = out_data + offset;
for (int k = 0; k < num; ++k) {
*dout_ptr = *din_ptr * diny_data;
dout_ptr++;
din_ptr++;
}
}
}
} else if (elt_type == "max") {
for (int i = 0; i < batch; ++i) {
for (int j = 0; j < channels; ++j) {
int offset = (i * channels + j) * num;
const dtype* din_ptr = x_data + offset;
const dtype diny_data = y_data[j];
dtype* dout_ptr = out_data + offset;
for (int k = 0; k < num; ++k) {
*dout_ptr = std::max(*din_ptr, diny_data);
dout_ptr++;
din_ptr++;
}
}
}
} else {
LOG(FATAL) << "unsupported Elementwise type: " << elt_type;
}
}
void test_elementwise_add(std::vector<int64_t> x_dims,
std::vector<int64_t> y_dims,
int axis) {
// prepare input&output variables
Scope scope;
std::string x_var_name = "x";
std::string y_var_name = "y";
std::string out_var_name = "out";
std::string out_ref_var_name = "out_ref";
auto* x = scope.Var(x_var_name)->GetMutable<Tensor>();
auto* y = scope.Var(y_var_name)->GetMutable<Tensor>();
auto* out = scope.Var(out_var_name)->GetMutable<Tensor>();
auto* out_ref = scope.Var(out_ref_var_name)->GetMutable<Tensor>();
x->Resize(x_dims);
if (y_dims.size() == 0) {
y->Resize(x_dims);
} else {
y->Resize(y_dims);
}
// initialize input&output data
FillTensor<float>(x);
FillTensor<float>(y);
// initialize op desc
cpp::OpDesc opdesc;
opdesc.SetType("elementwise_add");
opdesc.SetInput("X", {x_var_name});
opdesc.SetInput("Y", {y_var_name});
opdesc.SetOutput("Out", {out_var_name});
opdesc.SetAttr("axis", axis);
// create and convert op to XPU model, then run it on XPU
auto op = CreateOp<operators::ElementwiseOp>(opdesc, &scope);
LauchOp(op, {x_var_name, y_var_name}, {out_var_name});
out_ref->CopyDataFrom(*out);
// execute reference implementation and save to output tensor
elementwise_add_ref<float>(op);
// compare results
auto* out_data = out->mutable_data<float>();
auto* out_ref_data = out_ref->mutable_data<float>();
for (int i = 0; i < out->dims().production(); i++) {
EXPECT_NEAR(out_data[i], out_ref_data[i], 1e-5);
}
}
// xpu's bias_add only support y with one dimension
TEST(XPUBridges, elementwise_add) {
test_elementwise_add({1, 2, 3, 4}, {1}, 0);
test_elementwise_add({1, 2, 3, 4}, {2}, 1);
test_elementwise_add({2, 2, 3, 4}, {3}, 2);
test_elementwise_add({2, 2, 3, 4}, {4}, 3);
test_elementwise_add({2, 2, 3, 4}, {4}, -1);
test_elementwise_add({2, 2, 3, 4}, {}, -1);
}
} // namespace bridges
} // namespace xpu
} // namespace kernels
} // namespace lite
} // namespace paddle
USE_LITE_OP(elementwise_add);
USE_XPU_BRIDGE(elementwise_add);
lite/kernels/xpu/bridges/pool_op_test.cc 已删除 100644 → 0
浏览文件 @ a34f06a1
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// 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.
#include "lite/operators/pool_op.h"
#include <gtest/gtest.h>
#include "lite/core/op_registry.h"
#include "lite/kernels/xpu/bridges/registry.h"
#include "lite/kernels/xpu/bridges/test_helper.h"
namespace paddle {
namespace lite {
namespace kernels {
namespace xpu {
namespace bridges {
void pool_ref(const std::shared_ptr<operators::PoolOpLite> op) {
Scope* scope = op->scope();
const OpInfo* op_info = op->op_info();
auto x = scope->FindVar(op_info->Input("X").front())->GetMutable<Tensor>();
auto out =
scope->FindVar(op_info->Output("Out").front())->GetMutable<Tensor>();
auto& in_dims = x->dims();
auto& out_dims = out->dims();
const float* src_ptr = x->data<const float>();
float* dst_ptr = out->mutable_data<float>();
std::vector<int> ksize = op_info->GetAttr<std::vector<int>>("ksize");
std::vector<int> strides = op_info->GetAttr<std::vector<int>>("strides");
std::vector<int> paddings = op_info->GetAttr<std::vector<int>>("paddings");
bool exclusive = op_info->GetAttr<bool>("exclusive");
std::string pooling_type = op_info->GetAttr<std::string>("pooling_type");
bool global_pooling = op_info->GetAttr<bool>("global_pooling");
int in_n = in_dims[0];
int in_c = in_dims[1];
int in_h = in_dims[2];
int in_w = in_dims[3];
int size_in_n = in_c * in_h * in_w;
int size_in_c = in_h * in_w;
int out_h = out_dims[2];
int out_w = out_dims[3];
int size_out_n = in_c * out_h * out_w;
int size_out_c = out_h * out_w;
int window_h = ksize[0];
int window_w = ksize[1];
int stride_h = strides[0];
int stride_w = strides[1];
int pad_h = paddings[0];
int pad_w = paddings[2];
if (global_pooling == true) {
for (int n = 0; n < in_n; ++n) {
for (int c = 0; c < in_c; ++c) {
const float* src = src_ptr + n * size_in_n + c * size_in_c;
float res = src[0];
if (pooling_type == "max") {
for (int i = 1; i < size_in_c; ++i) {
float cur_val = src[i];
res = cur_val > res ? cur_val : res;
}
} else if (pooling_type == "avg") {
for (int i = 1; i < size_in_c; ++i) {
float cur_val = src[i];
res += cur_val;
}
res /= size_in_c;
}
dst_ptr[n * size_out_n + c] = res;
}
}
} else {
for (int n = 0; n < in_n; ++n) {
for (int c = 0; c < in_c; ++c) {
for (int h = 0; h < out_h; ++h) {
int sh = h * stride_h;
int eh = sh + window_h;
sh = (sh - pad_h) < 0 ? 0 : sh - pad_h;
eh = (eh - pad_h) > in_h ? in_h : eh - pad_h;
for (int w = 0; w < out_w; ++w) {
int sw = w * stride_w;
int ew = sw + window_w;
sw = (sw - pad_w) < 0 ? 0 : sw - pad_w;
ew = (ew - pad_w) > in_w ? in_w : ew - pad_w;
int pooling_size = (ew - sw) * (eh - sh);
if (pooling_size == 0) continue;
float res = 0.f;
for (int kh = sh; kh < eh; ++kh) {
for (int kw = sw; kw < ew; ++kw) {
int src_idx = n * size_in_n + c * size_in_c + kh * in_w + kw;
if (kh == sh && kw == sw) {
res = src_ptr[src_idx];
} else {
if (pooling_type == "max") {
res = res >= src_ptr[src_idx] ? res : src_ptr[src_idx];
}
if (pooling_type == "avg") {
res += src_ptr[src_idx];
}
}
}
}
if (pooling_type == "avg") {
if (exclusive) {
res /= pooling_size;
} else {
res /= window_h * window_w;
}
}
dst_ptr[n * size_out_n + c * size_out_c + h * out_w + w] = res;
}
}
}
}
}
}
void test_pool(int bs,
int ic,
int ih,
int iw,
std::string pooling_type,
bool ceil_mode,
bool global_pooling,
bool exclusive,
int ksize,
int stride,
int padding) {
// prepare input&output variables
Scope scope;
std::string x_var_name = "x";
std::string out_var_name = "out";
std::string out_ref_var_name = "out_ref";
auto* x = scope.Var(x_var_name)->GetMutable<Tensor>();
auto* out = scope.Var(out_var_name)->GetMutable<Tensor>();
auto* out_ref = scope.Var(out_ref_var_name)->GetMutable<Tensor>();
x->Resize({bs, ic, ih, iw});
// initialize input&output data
FillTensor<float>(x);
// initialize op desc
cpp::OpDesc opdesc;
opdesc.SetType("pool2d");
opdesc.SetInput("X", {x_var_name});
opdesc.SetOutput("Out", {out_var_name});
opdesc.SetAttr("pooling_type", pooling_type);
opdesc.SetAttr("ksize", std::vector<int>({ksize, ksize}));
opdesc.SetAttr("global_pooling", global_pooling);
opdesc.SetAttr("exclusive", exclusive);
opdesc.SetAttr("strides", std::vector<int>({stride, stride}));
opdesc.SetAttr("paddings",
std::vector<int>({padding, padding, padding, padding}));
opdesc.SetAttr("ceil_mode", ceil_mode);
// create and convert op to XPU model, then run it on XPU
auto op = CreateOp<operators::PoolOpLite>(opdesc, &scope);
LauchOp(op, {x_var_name}, {out_var_name});
out_ref->CopyDataFrom(*out);
// execute reference implementation and save to output tensor
pool_ref(op);
// compare results
auto* out_data = out->mutable_data<float>();
auto* out_ref_data = out_ref->mutable_data<float>();
for (int i = 0; i < out->dims().production(); i++) {
EXPECT_NEAR(out_data[i], out_ref_data[i], 1e-5);
}
}
TEST(XPUBridges, pool) {
for (auto pooling_type : {"max", "avg"}) {
for (auto bs : {1, 3}) {
for (auto ic : {2}) {
for (auto ih : {3}) {
for (auto iw : {4}) {
test_pool(bs, ic, ih, iw, pooling_type, true, true, true, 0, 1, 0);
}
}
}
}
}
for (auto pooling_type : {"max"}) {
for (auto ceil_mode : {true, false}) {
for (auto ksize : {2, 3}) {
for (auto stride : {1, 2}) {
for (auto padding : {0, 1}) {
for (auto bs : {1, 3}) {
for (auto ic : {2}) {
for (auto ih : {3}) {
for (auto iw : {4}) {
test_pool(bs,
ic,
ih,
iw,
pooling_type,
ceil_mode,
false,
true,
ksize,
stride,
padding);
}
}
}
}
}
}
}
}
}
for (auto pooling_type : {"avg"}) {
for (auto ceil_mode : {true, false}) {
for (auto exclusive : {true, false}) {
for (auto ksize : {2, 3}) {
for (auto stride : {1, 2}) {
for (auto padding : {0, 1}) {
for (auto bs : {1, 3}) {
for (auto ic : {2}) {
for (auto ih : {3}) {
for (auto iw : {4}) {
test_pool(bs,
ic,
ih,
iw,
pooling_type,
ceil_mode,
false,
exclusive,
ksize,
stride,
padding);
}
}
}
}
}
}
}
}
}
}
}
} // namespace bridges
} // namespace xpu
} // namespace kernels
} // namespace lite
} // namespace paddle
USE_LITE_OP(pool2d);
USE_XPU_BRIDGE(pool2d);
lite/tests/kernels/elementwise_compute_test.cc
浏览文件 @ 5dac279a
...@@ -234,7 +234,7 @@ TEST(Elementwise, precision) { ...@@ -234,7 +234,7 @@ TEST(Elementwise, precision) {
return; return;
#endif #endif
// TestEltDims(place, abs_error); TestEltDims(place, abs_error);
TestEltTypes(place, abs_error); TestEltTypes(place, abs_error);
TestEltFuseAct(place, abs_error); TestEltFuseAct(place, abs_error);
} }
......
lite/tests/kernels/pool_compute_test.cc
浏览文件 @ 5dac279a
...@@ -352,6 +352,8 @@ TEST(Pool, precision) { ...@@ -352,6 +352,8 @@ TEST(Pool, precision) {
#if defined(LITE_WITH_NPU) #if defined(LITE_WITH_NPU)
place = TARGET(kNPU); place = TARGET(kNPU);
abs_error = 1e-2; // Using fp16 in NPU abs_error = 1e-2; // Using fp16 in NPU
#elif defined(LITE_WITH_XPU)
place = TARGET(kXPU);
#else #else
return; return;
#endif #endif
......
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