未验证 提交 604c5d4f 编写于 作者: Y yiicy 提交者: GitHub

[OPENCL] add instance norm kernel and ut, test=develop (#3122)

add instance norm kernel and ut
上级 1bd34c3f
/* Copyright (c) 2018 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 <cl_common.h>
__kernel void instance_norm(__read_only image2d_t input,
__write_only image2d_t output,
__read_only image2d_t scale,
__read_only image2d_t bias,
const float epsilon,
const int in_h,
const int in_w){
__local CL_DTYPE4 saved_mean[1024];
__local CL_DTYPE4 saved_variance[1024];
const int lid = get_local_id(0);
const int lsize = get_local_size(0);
const int gidx = get_group_id(0);
const int gidy = get_group_id(1);
const int spatial_size = in_h * in_w;
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST;
CL_DTYPE4 mean = (CL_DTYPE4)(0.f, 0.f, 0.f, 0.f);
CL_DTYPE4 variance = (CL_DTYPE4)(0.f, 0.f, 0.f, 0.f);
CL_DTYPE4 vepsilon = (CL_DTYPE4)(epsilon, epsilon, epsilon, epsilon);
const int x_offset = gidx * in_w;
const int y_offset = gidy * in_h;
int2 coor;
for (int i = lid; i < spatial_size; i += lsize) {
coor.x = i % in_w + x_offset;
coor.y = i / in_w + y_offset;
CL_DTYPE4 pixel = READ_IMG_TYPE(CL_DTYPE_CHAR, input, sampler, coor);
mean += pixel;
variance += pixel * pixel;
}
saved_mean[lid] = mean;
saved_variance[lid] = variance;
barrier(CLK_LOCAL_MEM_FENCE);
//! do reduction
int dynamic_size = lsize >> 1;
for (; dynamic_size > 0; dynamic_size >>= 1){
if (lid < dynamic_size) {
saved_mean[lid] += saved_mean[lid + dynamic_size];
saved_variance[lid] += saved_variance[lid + dynamic_size];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
mean = saved_mean[0] / spatial_size;
variance = saved_variance[0] / spatial_size - mean * mean;
variance = rsqrt(variance + vepsilon);
//! do instance norm
coor.x = gidx;
coor.y = gidy;
CL_DTYPE4 vscale = READ_IMG_TYPE(CL_DTYPE_CHAR, scale, sampler, coor);
vscale *= variance;
CL_DTYPE4 vbias = READ_IMG_TYPE(CL_DTYPE_CHAR, bias, sampler, coor);
for (int i = lid; i < spatial_size; i += lsize) {
coor.x = i % in_w + x_offset;
coor.y = i / in_w + y_offset;
CL_DTYPE4 pixel = READ_IMG_TYPE(CL_DTYPE_CHAR, input, sampler, coor);
pixel = (pixel - mean) * vscale + vbias;
WRITE_IMG_TYPE(CL_DTYPE_CHAR, output, coor, pixel);
}
}
......@@ -29,6 +29,7 @@ add_kernel(scale_opencl OPENCL basic SRCS scale_image_compute.cc DEPS ${cl_kerne
add_kernel(grid_sampler_opencl OPENCL basic SRCS grid_sampler_image_compute.cc DEPS ${cl_kernel_deps})
add_kernel(lrn_opencl OPENCL basic SRCS lrn_image_compute.cc DEPS ${cl_kernel_deps})
add_kernel(bilinear_interp_opencl OPENCL basic SRCS bilinear_interp_image_compute.cc DEPS ${cl_kernel_deps})
add_kernel(instance_norm_opencl OPENCL basic SRCS instance_norm_image_compute.cc DEPS ${cl_kernel_deps})
# extra
# wait to add ...
......@@ -84,7 +85,9 @@ lite_cc_test(test_lrn_image_opencl SRCS lrn_image_compute_test.cc
lite_cc_test(test_bilinear_interp_image_opencl SRCS bilinear_interp_image_compute_test.cc
DEPS bilinear_interp_opencl op_registry program context)
lite_cc_test(test_instance_norm_image_opencl SRCS instance_norm_image_compute_test.cc
DEPS instance_norm_opencl op_registry program context)
######################
# buffer kernel #
######################
......
// 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 <memory>
#include <string>
#include "lite/backends/opencl/cl_half.h"
#include "lite/backends/opencl/cl_image_converter.h"
#include "lite/backends/opencl/cl_include.h"
#include "lite/core/kernel.h"
#include "lite/core/op_registry.h"
#include "lite/kernels/opencl/image_helper.h"
#include "lite/operators/op_params.h"
#include "lite/utils/logging.h"
#include "lite/utils/replace_stl/stream.h"
namespace paddle {
namespace lite {
namespace kernels {
namespace opencl {
class InstanceNormImageCompute : public KernelLite<TARGET(kOpenCL),
PRECISION(kFP16),
DATALAYOUT(kImageDefault)> {
public:
using param_t = operators::InstanceNormParam;
std::string doc() const override {
return "InstanceNorm using cl::Image2D(ImageDefault/RGBA), kFP16";
}
void PrepareForRun() override {
instance_norm_param_ = param_.get_mutable<param_t>();
auto channel = instance_norm_param_->scale->dims()[0];
auto batch = instance_norm_param_->x->dims()[0];
int64_t cgroup = (channel + 3) / 4;
int64_t cround = cgroup * 4;
std::vector<half_t> scale_img(cround * batch);
std::vector<half_t> bias_img(cround * batch);
const float* scale_data = instance_norm_param_->scale->data<float>();
const float* bias_data = instance_norm_param_->bias->data<float>();
//! init scale_img bias_img data
for (int i = 0; i < channel; ++i) {
scale_img[i] = Float2Half(scale_data[i]);
bias_img[i] = Float2Half(bias_data[i]);
}
for (int i = channel; i < cround; ++i) {
scale_img[i] = Float2Half(0.f);
bias_img[i] = Float2Half(0.f);
}
for (int i = 1; i < batch; ++i) {
memcpy(scale_img.data() + i * cround,
scale_img.data(),
cround * sizeof(half_t));
memcpy(bias_img.data() + i * cround,
bias_img.data(),
cround * sizeof(half_t));
}
DDim scale_img_size{{cgroup, batch}};
scale_image_.mutable_data<half_t, cl::Image2D>(
scale_img_size[0], scale_img_size[1], scale_img.data());
bias_image_.mutable_data<half_t, cl::Image2D>(
scale_img_size[0], scale_img_size[1], bias_img.data());
auto& context = ctx_->As<OpenCLContext>();
context.cl_context()->AddKernel(
kernel_func_name_, "image/instance_norm_kernel.cl", build_options_);
VLOG(1) << "kernel_func_name_:" << kernel_func_name_;
}
void Run() override {
auto& context = ctx_->As<OpenCLContext>();
CHECK(context.cl_context() != nullptr);
auto* x = instance_norm_param_->x;
auto* out = instance_norm_param_->out;
auto in_dims = x->dims();
int batch = in_dims[0];
int channel = in_dims[1];
int in_h = in_dims[2];
int in_w = in_dims[3];
VLOG(4) << "x->target():" << TargetToStr(x->target());
VLOG(4) << "out->target():" << TargetToStr(out->target());
VLOG(4) << "x->dims():" << in_dims;
auto out_image_shape = InitImageDimInfoWith(in_dims);
auto* x_img = x->data<half_t, cl::Image2D>();
auto* out_img = out->mutable_data<half_t, cl::Image2D>(
out_image_shape["width"], out_image_shape["height"]);
VLOG(4) << "out_image_shape[w,h]: " << out_image_shape["width"] << " "
<< out_image_shape["height"];
VLOG(4) << "in_h: " << in_h << ", in_w: " << in_w;
int threads = 512;
int group_size_x = (channel + 3) / 4;
int group_size_y = batch;
auto local_work_size = cl::NDRange{static_cast<cl::size_type>(threads),
static_cast<cl::size_type>(1),
static_cast<cl::size_type>(1)};
auto global_work_size =
cl::NDRange{static_cast<cl::size_type>(group_size_x * threads),
static_cast<cl::size_type>(group_size_y),
static_cast<cl::size_type>(1)};
VLOG(4) << "local_work_size:[2D]:" << local_work_size[0] << " "
<< local_work_size[1] << " " << local_work_size[2];
VLOG(4) << "global_work_size:[2D]:" << global_work_size[0] << " "
<< global_work_size[1] << " " << global_work_size[2];
STL::stringstream kernel_key;
kernel_key << kernel_func_name_ << build_options_;
auto kernel = context.cl_context()->GetKernel(kernel_key.str());
auto* scale_img = scale_image_.data<half_t, cl::Image2D>();
auto* bias_img = bias_image_.data<half_t, cl::Image2D>();
float epsilon = instance_norm_param_->epsilon;
int arg_idx = 0;
cl_int status = kernel.setArg(arg_idx++, *x_img);
CL_CHECK_FATAL(status);
status = kernel.setArg(arg_idx++, *out_img);
CL_CHECK_FATAL(status);
status = kernel.setArg(arg_idx++, *scale_img);
CL_CHECK_FATAL(status);
status = kernel.setArg(arg_idx++, *bias_img);
CL_CHECK_FATAL(status);
status = kernel.setArg(arg_idx++, epsilon);
CL_CHECK_FATAL(status);
status = kernel.setArg(arg_idx++, in_h);
CL_CHECK_FATAL(status);
status = kernel.setArg(arg_idx++, in_w);
CL_CHECK_FATAL(status);
status = context.cl_context()->GetCommandQueue().enqueueNDRangeKernel(
kernel,
cl::NullRange,
global_work_size,
local_work_size,
nullptr,
event_.get());
CL_CHECK_FATAL(status);
context.cl_wait_list()->emplace(out_img, event_);
}
protected:
param_t* instance_norm_param_{nullptr};
std::string kernel_func_name_{"instance_norm"};
std::string build_options_{"-DCL_DTYPE_half"};
std::shared_ptr<cl::Event> event_{new cl::Event};
Tensor scale_image_;
Tensor bias_image_;
};
} // namespace opencl
} // namespace kernels
} // namespace lite
} // namespace paddle
namespace ocl = paddle::lite::kernels::opencl;
REGISTER_LITE_KERNEL(instance_norm,
kOpenCL,
kFP16,
kImageDefault,
ocl::InstanceNormImageCompute,
ImageDefault)
.BindInput("X",
{LiteType::GetTensorTy(TARGET(kOpenCL),
PRECISION(kFP16),
DATALAYOUT(kImageDefault))})
.BindOutput("Y",
{LiteType::GetTensorTy(TARGET(kOpenCL),
PRECISION(kFP16),
DATALAYOUT(kImageDefault))})
.BindInput("Scale", {LiteType::GetTensorTy(TARGET(kARM))})
.BindInput("Bias", {LiteType::GetTensorTy(TARGET(kARM))})
.BindOutput("SavedMean", {LiteType::GetTensorTy(TARGET(kARM))})
.BindOutput("SavedVariance", {LiteType::GetTensorTy(TARGET(kARM))})
.Finalize();
// 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 <gflags/gflags.h>
#include <gtest/gtest.h>
#include <memory>
#include <random>
#include "lite/backends/opencl/target_wrapper.h"
#include "lite/core/op_registry.h"
#include "lite/core/profile/timer.h"
#include "lite/core/tensor.h"
#include "lite/kernels/opencl/test_helper.h"
#define FP16_MAX_DIFF (5e-3)
DEFINE_int32(warmup, 0, "warmup times");
DEFINE_int32(repeats, 1, "repeats times");
using paddle::lite::profile::Timer;
namespace paddle {
namespace lite {
void instance_norm_ref(Tensor* x,
Tensor* y,
Tensor* scale,
Tensor* bias,
Tensor* saved_mean,
Tensor* saved_variance,
float epsilon) {
auto x_data = x->data<float>();
auto scale_data = scale->data<float>();
auto bias_data = bias->data<float>();
auto y_data = y->mutable_data<float>();
auto saved_mean_data = saved_mean->mutable_data<float>();
auto saved_variance_data = saved_variance->mutable_data<float>();
int n = x->dims()[0];
int c = x->dims()[1];
int spatial_size = x->dims()[2] * x->dims()[3];
// compute mean
for (int i = 0; i < n * c; ++i) {
const float* x_ptr = x_data + i * spatial_size;
float sum = 0.f;
for (int j = 0; j < spatial_size; ++j) {
sum += x_ptr[j];
}
saved_mean_data[i] = sum / spatial_size;
}
// compute variance
for (int i = 0; i < n * c; ++i) {
const float* x_ptr = x_data + i * spatial_size;
float sum = 0.f;
for (int j = 0; j < spatial_size; ++j) {
sum += (x_ptr[j] - saved_mean_data[i]) * (x_ptr[j] - saved_mean_data[i]);
}
saved_variance_data[i] = 1.f / sqrtf(sum / spatial_size + epsilon);
}
// compute out
for (int i = 0; i < n * c; ++i) {
const float* x_ptr = x_data + i * spatial_size;
float* y_ptr = y_data + i * spatial_size;
float scale_val = scale_data[i % c];
float bias_val = bias_data[i % c];
for (int j = 0; j < spatial_size; ++j) {
y_ptr[j] =
scale_val * (x_ptr[j] - saved_mean_data[i]) * saved_variance_data[i] +
bias_val;
}
}
}
// #define INSTANCE_NORM_FP16_LOOP_TEST
// #define INSTANCE_NORM_FP16_PRINT_RESULT
TEST(instance_norm_image2d, compute) {
#ifdef INSTANCE_NORM_FP16_LOOP_TEST
for (auto n : {1, 3}) {
for (auto c : {1, 3, 8, 32, 65}) {
for (auto h : {4, 20, 64, 112, 224}) {
for (auto w : {2, 20, 64, 112, 224}) {
#else
const int n = 1;
const int c = 32;
const int h = 224;
const int w = 224;
#endif // INSTANCE_NORM_FP16_LOOP_TEST
LOG(INFO) << "======== input shape[n,c,h,w]:" << n << " " << c << " "
<< h << " " << w << " ========";
auto kernels =
KernelRegistry::Global().Create("instance_norm",
TARGET(kOpenCL),
PRECISION(kFP16),
DATALAYOUT(kImageDefault));
ASSERT_FALSE(kernels.empty());
auto kernel = std::move(kernels.front());
LOG(INFO) << "get kernel:" << kernel->doc();
lite::Tensor x, out, out_ref, scale, bias, saved_mean, saved_variance;
operators::InstanceNormParam param;
param.x = &x;
param.out = &out;
param.scale = &scale;
param.bias = &bias;
param.saved_mean = &saved_mean;
param.saved_variance = &saved_variance;
param.epsilon = 1e-5;
std::unique_ptr<KernelContext> context(new KernelContext);
context->As<OpenCLContext>().InitOnce();
kernel->SetParam(param);
std::unique_ptr<KernelContext> instance_context(new KernelContext);
context->As<OpenCLContext>().CopySharedTo(
&(instance_context->As<OpenCLContext>()));
kernel->SetContext(std::move(instance_context));
const DDim in_dim = DDim(std::vector<DDim::value_type>{n, c, h, w});
x.Resize(in_dim);
out.Resize(in_dim);
out_ref.Resize(in_dim);
scale.Resize({c});
bias.Resize({c});
saved_mean.Resize({n * c});
saved_variance.Resize({n * c});
auto* x_data = x.mutable_data<float>();
auto* scale_data = scale.mutable_data<float>();
auto* bias_data = bias.mutable_data<float>();
auto* saved_mean_data = saved_mean.mutable_data<float>();
auto* saved_variance_data = saved_variance.mutable_data<float>();
std::default_random_engine engine;
std::uniform_real_distribution<float> dist(-1, 1);
int sum = n * c * h * w;
for (int i = 0; i < sum; ++i) {
x_data[i] = dist(engine);
}
for (int i = 0; i < c; ++i) {
scale_data[i] = dist(engine);
bias_data[i] = dist(engine);
}
//! run reference instance norm
instance_norm_ref(
&x, &out_ref, &scale, &bias, &saved_mean, &saved_variance, 1e-5);
LOG(INFO) << "prepare input";
CLImageConverterDefault* default_converter =
new CLImageConverterDefault();
DDim x_image_shape = default_converter->InitImageDimInfoWith(in_dim);
LOG(INFO) << "x_image_shape = " << x_image_shape[0] << " "
<< x_image_shape[1];
std::vector<half_t> x_image_data(x_image_shape.production() *
4); // 4 : RGBA
default_converter->NCHWToImage(x_data, x_image_data.data(), in_dim);
auto* x_image = x.mutable_data<half_t, cl::Image2D>(
x_image_shape[0], x_image_shape[1], x_image_data.data());
auto* out_image = out.mutable_data<half_t, cl::Image2D>(
x_image_shape[0], x_image_shape[1]);
//! warm up
for (int i = 0; i < FLAGS_warmup; ++i) {
kernel->Launch();
}
context->As<OpenCLContext>().cl_context()->GetCommandQueue().finish();
//! compute
Timer t0;
t0.Start();
for (int i = 0; i < FLAGS_repeats; ++i) {
kernel->Launch();
}
context->As<OpenCLContext>().cl_context()->GetCommandQueue().finish();
t0.Stop();
double gops = 6 * sum;
LOG(INFO) << "avg time: " << t0.LapTimes().Avg() / FLAGS_repeats
<< " ms, "
<< "avg GOPs: "
<< 1e-6 * gops * FLAGS_repeats / t0.LapTimes().Avg()
<< " GOPs";
const size_t cl_image2d_row_pitch{0};
const size_t cl_image2d_slice_pitch{0};
half_t* out_image_data = new half_t[x_image_shape.production() * 4];
TargetWrapperCL::ImgcpySync(out_image_data,
out_image,
x_image_shape[0],
x_image_shape[1],
cl_image2d_row_pitch,
cl_image2d_slice_pitch,
IoDirection::DtoH);
float* out_data = new float[x_image_shape.production() * 4];
default_converter->ImageToNCHW(
out_image_data, out_data, x_image_shape, in_dim);
// result
#ifdef INSTANCE_NORM_FP16_PRINT_RESULT
LOG(INFO) << "---- print kernel result (input -> output) ----";
for (int eidx = 0; eidx < in_dim.production(); ++eidx) {
std::cout << x_data[eidx] << " -> " << out_data[eidx] << std::endl;
}
#endif // INSTANCE_NORM_FP16_PRINT_RESULT
auto* out_ref_data = out_ref.data<float>();
for (int i = 0; i < in_dim.production(); i++) {
auto abs_diff = abs(out_data[i] - out_ref_data[i]);
auto relative_diff =
COMPUTE_RELATIVE_DIFF(out_data[i], out_ref_data[i]);
EXPECT_EQ(
(relative_diff <= FP16_MAX_DIFF) || (abs_diff <= FP16_MAX_DIFF),
true);
if ((relative_diff > FP16_MAX_DIFF) && (abs_diff > FP16_MAX_DIFF)) {
LOG(ERROR) << "error idx:" << i << ", in_data[" << i
<< "]: " << x_data[i] << ", out_data[" << i
<< "]: " << out_data[i] << ", out_ref[" << i
<< "]: " << out_ref_data[i]
<< ", abs_diff: " << abs_diff
<< ", relative_diff: " << relative_diff
<< ", FP16_MAX_DIFF: " << FP16_MAX_DIFF;
}
}
delete[] out_data;
delete[] out_image_data;
#ifdef INSTANCE_NORM_FP16_LOOP_TEST
} // w
} // h
} // c
} // n
#else
// nothing to do.
#endif
}
} // namespace lite
} // namespace paddle
USE_LITE_KERNEL(instance_norm, kOpenCL, kFP16, kImageDefault, ImageDefault);
Markdown is supported
0% .
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