Skip to content

Commit

Permalink
[Cherry-pick]FFT function enhancements and bugfixes (PaddlePaddle#36537)
Browse files Browse the repository at this point in the history 
* update fft api path (PaddlePaddle#36219)

* update fft api path
* add sample code for ihfft2

Co-authored-by: chenfeiyu <chenfeiyu@baidu.com>

* fix fft axis (PaddlePaddle#36321)

fix: `-1` is used when fft's axis is `0`

* use unified external error message for cufft api (PaddlePaddle#36114)

* fft: modify sample code result (PaddlePaddle#36325)

* dynamic load mkl as a fft backend when it is avaialble and requested (PaddlePaddle#36414)

* add rocm support for fft api (PaddlePaddle#36415)

* move signal apis

* move fft and signal API path (#2)

* move signal apis

* move fft.py and signal.py to paddle/, fix typos

* fix relative imports from fft.py and signal.py

* fix typos in signal.py (#3)

* move signal apis

* move fft.py and signal.py to paddle/, fix typos

* fix relative imports from fft.py and signal.py

* fix typos

* disable Cache when CUFFT_VERSION >= 10200 (#4)

* move signal apis

* move fft.py and signal.py to paddle/, fix typos

* fix relative imports from fft.py and signal.py

* fix typos

* Add LRUCache for fft plans

* add LRUCache for cuff and hipfft (#5)

* move signal apis

* move fft.py and signal.py to paddle/, fix typos

* fix relative imports from fft.py and signal.py

* fix typos

* WIP: add cache

* delete move constructor and operator= for CuFFTHandle and FFTConfig

* remove log from CuFFTHandle and FFTConfig

* add lrucache for fft rocm backend

* disable LRUCache when CUFFT_VERSION >= 10200

* disbale copy and move for hipFFTHandle; format code

Co-authored-by: Xiaoxu Chen <chenxx_id@163.com>

* remove debug message of cufftHandler

* roll_op: support Tensor as input for shifts (PaddlePaddle#36727)

* fix fftshift/ifftshift on static mode

* update roll_op version

* add more test cases for fftshift/ifftshift

Co-authored-by: zhiboniu <31800336+zhiboniu@users.noreply.github.com>
Co-authored-by: chenfeiyu <chenfeiyu@baidu.com>
Co-authored-by: LJQ❤️ <33169170+lijiaqi0612@users.noreply.github.com>
  • Loading branch information
@cxxly @zhiboniu @lijiaqi0612
4 people committed Oct 28, 2021
1 parent 96edcea commit 11b9f5f
Show file tree
Hide file tree
Showing 29 changed files with 1,413 additions and 549 deletions.
4 changes: 2 additions & 2 deletions cmake/third_party.cmake
View file
Original file line number Diff line number Diff line change
Expand Up @@ -255,8 +255,8 @@ if(WITH_GPU)
include(external/cub) # download cub
list(APPEND third_party_deps extern_cub)
endif()
set(URL "https://paddlepaddledeps.bj.bcebos.com/externalErrorMsg.tar.gz" CACHE STRING "" FORCE)
file_download_and_uncompress(${URL} "externalError" MD5 061f3b7895aadcbe2c3ed592590f8b10) # download file externalErrorMsg.tar.gz
set(URL "https://paddlepaddledeps.bj.bcebos.com/externalErrorMsg_20210928.tar.gz" CACHE STRING "" FORCE)
file_download_and_uncompress(${URL} "externalError" MD5 a712a49384e77ca216ad866712f7cafa) # download file externalErrorMsg.tar.gz
if(WITH_TESTING)
# copy externalErrorMsg.pb, just for unittest can get error message correctly.
set(SRC_DIR ${THIRD_PARTY_PATH}/externalError/data)
Expand Down
16 changes: 13 additions & 3 deletions paddle/fluid/operators/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -105,10 +105,20 @@ else()
op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
endif()

if (WITH_GPU AND (NOT WITH_ROCM))
op_library(spectral_op SRCS spectral_op.cc spectral_op.cu DEPS dynload_cuda ${OP_HEADER_DEPS})
if (WITH_GPU OR WITH_ROCM)
if (MKL_FOUND AND WITH_ONEMKL)
op_library(spectral_op SRCS spectral_op.cc spectral_op.cu DEPS dynload_cuda dynload_mklrt ${OP_HEADER_DEPS})
target_include_directories(spectral_op PRIVATE ${MKL_INCLUDE})
else()
op_library(spectral_op SRCS spectral_op.cc spectral_op.cu DEPS dynload_cuda ${OP_HEADER_DEPS})
endif()
else()
op_library(spectral_op SRCS spectral_op.cc DEPS ${OP_HEADER_DEPS})
if (MKL_FOUND AND WITH_ONEMKL)
op_library(spectral_op SRCS spectral_op.cc DEPS dynload_mklrt ${OP_HEADER_DEPS})
target_include_directories(spectral_op PRIVATE ${MKL_INCLUDE})
else()
op_library(spectral_op SRCS spectral_op.cc DEPS ${OP_HEADER_DEPS})
endif()
endif()

op_library(lstm_op DEPS ${OP_HEADER_DEPS} lstm_compute)
Expand Down
52 changes: 33 additions & 19 deletions paddle/fluid/operators/roll_op.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -40,21 +40,23 @@ class RollOp : public framework::OperatorWithKernel {
auto dims = ctx->Attrs().Get<std::vector<int64_t>>("axis");
auto shifts = ctx->Attrs().Get<std::vector<int64_t>>("shifts");

if (dims.size() != 0) {
PADDLE_ENFORCE_EQ(dims.size(), shifts.size(),
platform::errors::InvalidArgument(
"When dims.size() != 0, dims.size() "
"should be equal to "
"shifts.size(). But received "
"dims.size() = %d, shifts.size() = %d",
dims.size(), shifts.size()));
} else {
PADDLE_ENFORCE_EQ(shifts.size(), 1,
platform::errors::InvalidArgument(
"When dims.size() == 0, shifts.size() "
"should be equal to 1, But received "
"shifts.size() = %d",
shifts.size()));
if (!ctx->HasInput("ShiftsTensor")) {
if (dims.size() != 0) {
PADDLE_ENFORCE_EQ(dims.size(), shifts.size(),
platform::errors::InvalidArgument(
"When dims.size() != 0, dims.size() "
"should be equal to "
"shifts.size(). But received "
"dims.size() = %d, shifts.size() = %d",
dims.size(), shifts.size()));
} else {
PADDLE_ENFORCE_EQ(shifts.size(), 1,
platform::errors::InvalidArgument(
"When dims.size() == 0, shifts.size() "
"should be equal to 1, But received "
"shifts.size() = %d",
shifts.size()));
}
}

ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
Expand Down Expand Up @@ -105,6 +107,10 @@ class RollOpMaker : public framework::OpProtoAndCheckerMaker {
"The number of places by which the elements "
"of the tensor are shifted.")
.SetDefault({});
AddInput("ShiftsTensor",
"The number of places by which the elements of the tensor "
"are shifted.")
.AsDispensable();
AddAttr<std::vector<int64_t>>(
"axis",
"Axis along which to roll. It must have the same size "
Expand All @@ -129,6 +135,9 @@ class RollGradMaker : public framework::SingleGradOpMaker<T> {
void Apply(GradOpPtr<T> op) const override {
op->SetType("roll_grad");
op->SetInput("X", this->Input("X"));
if (this->HasInput("ShiftsTensor")) {
op->SetInput("ShiftsTensor", this->Input("ShiftsTensor"));
}
op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
op->SetAttrMap(this->Attrs());
Expand Down Expand Up @@ -174,7 +183,12 @@ REGISTER_OP_VERSION(roll)
"(std::vector<int64_t>) Axis along which to roll. "
"It must have the same size with shifts, or size = 0.",
std::vector<int64_t>())
.DeleteAttr(
"dims",
"(std::vector<int64_t>) Dims along which to roll. "
"It must have the same size with shifts, or size = 0."));
.DeleteAttr("dims",
"(std::vector<int64_t>) Dims along which to roll. "
"It must have the same size with shifts, or size = 0."))
.AddCheckpoint(
R"ROC(Upgrade roll add a dispensable input "ShiftsTensor".)ROC",
paddle::framework::compatible::OpVersionDesc().NewInput(
"ShiftsTensor",
"The number of places by which the elements of"
"the tensor are shifted."));
20 changes: 20 additions & 0 deletions paddle/fluid/operators/roll_op.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -59,6 +59,16 @@ class RollKernel<platform::CUDADeviceContext, T>
auto* in = context.Input<LoDTensor>("X");
auto* out = context.Output<LoDTensor>("Out");
std::vector<int64_t> shifts = context.Attr<std::vector<int64_t>>("shifts");
if (context.HasInput("ShiftsTensor")) {
const auto* shifts_tensor =
context.Input<framework::Tensor>("ShiftsTensor");
PADDLE_ENFORCE_EQ(
shifts_tensor->dims().size(), 1,
platform::errors::InvalidArgument(
"The rank of ShiftsTensor is expected to be 1, got %s",
shifts_tensor->dims().size()));
shifts = GetDataFromTensor<int64_t>(shifts_tensor);
}
std::vector<int64_t> dims = context.Attr<std::vector<int64_t>>("axis");

auto* in_data = in->data<T>();
Expand Down Expand Up @@ -134,6 +144,16 @@ class RollGradKernel<platform::CUDADeviceContext, T>
auto* in = context.Input<LoDTensor>(framework::GradVarName("Out"));
auto* out = context.Output<LoDTensor>(framework::GradVarName("X"));
std::vector<int64_t> shifts = context.Attr<std::vector<int64_t>>("shifts");
if (context.HasInput("ShiftsTensor")) {
const auto* shifts_tensor =
context.Input<framework::Tensor>("ShiftsTensor");
PADDLE_ENFORCE_EQ(
shifts_tensor->dims().size(), 1,
platform::errors::InvalidArgument(
"The rank of ShiftsTensor is expected to be 1, got %s",
shifts_tensor->dims().size()));
shifts = GetDataFromTensor<int64_t>(shifts_tensor);
}
std::vector<int64_t> dims = context.Attr<std::vector<int64_t>>("axis");

auto* in_data = in->data<T>();
Expand Down
17 changes: 17 additions & 0 deletions paddle/fluid/operators/roll_op.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,8 @@
#include <memory>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/utils.h"
#include "paddle/fluid/platform/enforce.h"

namespace paddle {
namespace operators {
Expand Down Expand Up @@ -85,6 +87,16 @@ class RollKernel : public framework::OpKernel<T> {
auto& input = input_var->Get<LoDTensor>();
auto* output = output_var->GetMutable<LoDTensor>();
std::vector<int64_t> shifts = context.Attr<std::vector<int64_t>>("shifts");
if (context.HasInput("ShiftsTensor")) {
const auto* shifts_tensor =
context.Input<framework::Tensor>("ShiftsTensor");
PADDLE_ENFORCE_EQ(
shifts_tensor->dims().size(), 1,
platform::errors::InvalidArgument(
"The rank of ShiftsTensor is expected to be 1, got %s",
shifts_tensor->dims().size()));
shifts = GetDataFromTensor<int64_t>(shifts_tensor);
}
std::vector<int64_t> dims = context.Attr<std::vector<int64_t>>("axis");

std::vector<T> out_vec;
Expand Down Expand Up @@ -123,6 +135,11 @@ class RollGradKernel : public framework::OpKernel<T> {
auto& input = input_var->Get<LoDTensor>();
auto* output = output_var->GetMutable<LoDTensor>();
std::vector<int64_t> shifts = context.Attr<std::vector<int64_t>>("shifts");
if (context.HasInput("ShiftsTensor")) {
const auto* shifts_tensor =
context.Input<framework::Tensor>("ShiftsTensor");
shifts = GetDataFromTensor<int64_t>(shifts_tensor);
}
std::vector<int64_t> dims = context.Attr<std::vector<int64_t>>("axis");

std::vector<T> out_vec;
Expand Down
Loading

0 comments on commit 11b9f5f

Please sign in to comment.