From a500dfa579907d8046e40a15e67558c350498976 Mon Sep 17 00:00:00 2001 From: sneaxiy Date: Tue, 18 Dec 2018 06:27:32 +0000 Subject: [PATCH] rewrite ddim test=develop --- paddle/fluid/framework/CMakeLists.txt | 2 +- paddle/fluid/framework/array.h | 74 ++- paddle/fluid/framework/ddim.cc | 303 ++++-------- paddle/fluid/framework/ddim.h | 148 ++++-- paddle/fluid/framework/dim.h | 441 ++++++------------ paddle/fluid/framework/dlpack_tensor.cc | 6 +- paddle/fluid/framework/dlpack_tensor.h | 2 +- paddle/fluid/framework/unroll_array_ops.h | 169 +++++++ .../fluid/operators/controlflow/logical_op.cc | 2 - paddle/fluid/operators/crop_op.h | 1 - paddle/fluid/operators/cudnn_lstm_op.cu.cc | 1 - .../fluid/operators/detail/strided_memcpy.h | 38 +- .../detection/generate_proposal_labels_op.cc | 2 - .../detection/generate_proposals_op.cc | 6 - .../detection/rpn_target_assign_op.cc | 1 - .../operators/elementwise/elementwise_op.h | 1 - paddle/fluid/operators/expand_op.h | 1 - paddle/fluid/operators/fc_op.cc | 1 - .../fused/fused_embedding_fc_lstm_op.cc | 18 +- paddle/fluid/operators/hinge_loss_op.cc | 1 - paddle/fluid/operators/log_loss_op.cc | 1 - .../fluid/operators/math/math_function_impl.h | 3 - paddle/fluid/operators/math/softmax_impl.h | 1 - .../fluid/operators/modified_huber_loss_op.cc | 1 - paddle/fluid/operators/mul_op.cc | 6 - paddle/fluid/operators/nce_op.cc | 1 - paddle/fluid/operators/norm_op.h | 1 - paddle/fluid/operators/psroi_pool_op.h | 1 - .../sequence_ops/sequence_slice_op.h | 2 - paddle/fluid/operators/strided_memcpy.h | 2 +- 30 files changed, 622 insertions(+), 615 deletions(-) create mode 100644 paddle/fluid/framework/unroll_array_ops.h diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 6d7a69c8c9e..023118d7407 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -36,7 +36,7 @@ add_subdirectory(details) proto_library(framework_proto SRCS framework.proto) proto_library(async_executor_proto SRCS data_feed.proto) -cc_library(ddim SRCS ddim.cc DEPS eigen3 boost) +cc_library(ddim SRCS ddim.cc DEPS eigen3 boost enforce) cc_test(ddim_test SRCS ddim_test.cc DEPS ddim) nv_test(dim_test SRCS dim_test.cu DEPS ddim) cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context) diff --git a/paddle/fluid/framework/array.h b/paddle/fluid/framework/array.h index be9efcd7492..aa0abc22a6b 100644 --- a/paddle/fluid/framework/array.h +++ b/paddle/fluid/framework/array.h @@ -15,34 +15,88 @@ #pragma once #include -#include "paddle/fluid/platform/hostdevice.h" +#include "paddle/fluid/framework/unroll_array_ops.h" +#include "paddle/fluid/platform/enforce.h" namespace paddle { namespace framework { + template class Array { - static_assert(N > 0, "The size of array must be larger than 0"); - public: - HOSTDEVICE Array() {} + static constexpr size_t kSize = N; - HOSTDEVICE explicit Array(const T &val) { - for (size_t i = 0; i < N; ++i) data_[i] = val; + HOSTDEVICE inline Array() = default; + + template + HOSTDEVICE inline explicit Array(const T &val, Args... args) { + UnrollVarArgsAssign::Run(data_, val, args...); } - HOSTDEVICE const T *Get() const { return data_; } + HOSTDEVICE inline void Fill(const T &val) { + UnrollFillConstant::Run(data_, val); + } - HOSTDEVICE T *GetMutable() { return data_; } + HOSTDEVICE inline const T *Get() const { return data_; } - HOSTDEVICE T &operator[](size_t index) { return data_[index]; } + HOSTDEVICE inline T *GetMutable() { return data_; } - HOSTDEVICE const T &operator[](size_t index) const { return data_[index]; } + HOSTDEVICE inline T &operator[](size_t index) { return data_[index]; } + + HOSTDEVICE inline const T &operator[](size_t index) const { + return data_[index]; + } HOSTDEVICE constexpr size_t size() const { return N; } + HOSTDEVICE inline bool operator==(const Array &other) const { + return UnrollCompare::Run(data_, other.data_); + } + + HOSTDEVICE inline bool operator!=(const Array &other) const { + return !(*this == other); + } + private: T data_[N]; }; +template +class Array { + public: + static constexpr size_t kSize = 0; + + HOSTDEVICE inline Array() = default; + + HOSTDEVICE inline void Fill(const T &val) {} + + HOSTDEVICE inline constexpr T *Get() const { return nullptr; } + + // Add constexpr to GetMutable() cause warning in MAC + HOSTDEVICE inline T *GetMutable() { return nullptr; } + + HOSTDEVICE inline T &operator[](size_t index) { +#ifndef __CUDA_ARCH__ + PADDLE_THROW("Array has no element"); +#endif + } + + HOSTDEVICE inline const T &operator[](size_t index) const { +#ifndef __CUDA_ARCH__ + PADDLE_THROW("Array has no element"); +#endif + } + + HOSTDEVICE constexpr size_t size() const { return 0; } + + HOSTDEVICE constexpr bool operator==(const Array &other) const { + return true; + } + + HOSTDEVICE constexpr bool operator!=(const Array &other) const { + return false; + } +}; + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ddim.cc b/paddle/fluid/framework/ddim.cc index 05e423b8a52..3640138e180 100644 --- a/paddle/fluid/framework/ddim.cc +++ b/paddle/fluid/framework/ddim.cc @@ -18,201 +18,131 @@ limitations under the License. */ namespace paddle { namespace framework { -/// @cond HIDDEN +template +struct DDimAssignFunctor { + static_assert(std::is_integral::value, "T must be integral type"); + using result_type = void; + explicit DDimAssignFunctor(const T* in) : in_(in) {} -template -Dim make_dim(const int64_t* d) { - return Dim(*d, make_dim(d + 1)); -} + template + inline void operator()(Dim& dim) { // NOLINT + UnrollAssign::Run(in_, dim.data()); + } + + const T* in_; +}; -template <> -Dim<0> make_dim<0>(const int64_t* d) { - return Dim<0>(*d); +DDim::DDim(const int* d, int n) : rank_(n) { + this->apply_visitor(DDimAssignFunctor(d)); } -void make_ddim(DDim& ddim, const int64_t* dims, int n) { - switch (n) { - case 0: - ddim = make_dim<0>(dims); - break; - case 1: - ddim = make_dim<1>(dims); - break; - case 2: - ddim = make_dim<2>(dims); - break; - case 3: - ddim = make_dim<3>(dims); - break; - case 4: - ddim = make_dim<4>(dims); - break; - case 5: - ddim = make_dim<5>(dims); - break; - case 6: - ddim = make_dim<6>(dims); - break; - case 7: - ddim = make_dim<7>(dims); - break; - case 8: - ddim = make_dim<8>(dims); - break; - case 9: - ddim = make_dim<9>(dims); - break; - default: - PADDLE_THROW("Dynamic dimensions must have between [1, 9] dimensions."); - } +DDim::DDim(const int64_t* d, int n) : rank_(n) { + this->apply_visitor(DDimAssignFunctor(d)); } -/// @endcond +template +Dim make_dim(const int64_t* d) { + Dim ret; + for (int i = 0; i < N; ++i) ret[i] = d[i]; + return ret; +} DDim make_ddim(std::initializer_list dims) { - DDim result(make_dim(0)); - make_ddim(result, dims.begin(), dims.size()); - return result; + return DDim(dims.begin(), dims.size()); } DDim make_ddim(const std::vector& dims) { - DDim result(make_dim(0)); - make_ddim(result, &dims[0], dims.size()); - return result; + return DDim(dims.data(), dims.size()); } DDim make_ddim(const std::vector& dims) { - std::vector res(dims.size()); - std::transform(dims.begin(), dims.end(), res.begin(), - [](int d) { return static_cast(d); }); - return make_ddim(res); + return DDim(dims.data(), dims.size()); } -/// @cond HIDDEN -// XXX For some reason, putting this in an anonymous namespace causes errors -class DynamicMutableIndexer : public boost::static_visitor { - public: - explicit DynamicMutableIndexer(int idx) : idx_(idx) {} - - template - int64_t& operator()(Dim& dim) const { - return dim[idx_]; - } - - private: - int idx_; -}; - -class DynamicConstIndexer : public boost::static_visitor { - public: - explicit DynamicConstIndexer(int idx) : idx_(idx) {} +struct DDimEqualityVisitor { + explicit DDimEqualityVisitor(const int64_t* d) : d_(d) {} template - int64_t operator()(const Dim& dim) const { - return dim[idx_]; + inline bool operator()(const Dim& self) const { + return UnrollCompare::Run(self.data(), d_); } - private: - int idx_; + const int64_t* d_; }; -/// @endcond - -int64_t& DDim::operator[](int idx) { - return boost::apply_visitor(DynamicMutableIndexer(idx), var); -} - -int64_t DDim::operator[](int idx) const { - return boost::apply_visitor(DynamicConstIndexer(idx), var); +bool DDim::operator==(const DDim& d) const { + return rank_ == d.rank_ && this->apply_visitor(DDimEqualityVisitor(d.data())); } -int DDim::size() const { return arity(*this); } - -bool DDim::operator==(DDim d) const { - if (var.which() != d.getVar().which()) { - return false; - } else { - std::vector v1 = vectorize(*this); - std::vector v2 = vectorize(d); +bool DDim::operator!=(const DDim& d) const { return !(*this == d); } - for (unsigned int i = 0; i < v1.size(); i++) { - if (v1[i] != v2[i]) { - return false; - } - } +struct DDimPlusVisitor { + explicit DDimPlusVisitor(const int64_t* d1, const int64_t* d2) + : d1_(d1), d2_(d2) {} - return true; + template + inline void operator()(Dim& self) const { + UnrollAdd::Run(d1_, d2_, self.data()); } -} - -bool DDim::operator!=(DDim d) const { return !(*this == d); } - -DDim DDim::operator+(DDim d) const { - std::vector v1 = vectorize(*this); - std::vector v2 = vectorize(d); - std::vector v3; - - assert(v1.size() == v2.size()); - - for (unsigned int i = 0; i < v1.size(); i++) { - v3.push_back(v1[i] + v2[i]); - } + const int64_t* d1_; + const int64_t* d2_; +}; - return make_ddim(v3); +DDim DDim::operator+(const DDim& d) const { + PADDLE_ENFORCE(rank_ == d.rank_); + DDim ret; + ret.rank_ = rank_; + ret.apply_visitor(DDimPlusVisitor(data(), d.data())); + return ret; } -DDim DDim::operator*(DDim d) const { - std::vector v1 = vectorize(*this); - std::vector v2 = vectorize(d); +struct DDimMulVisitor { + explicit DDimMulVisitor(const int64_t* d1, const int64_t* d2) + : d1_(d1), d2_(d2) {} - std::vector v3; - - assert(v1.size() == v2.size()); - - for (unsigned int i = 0; i < v1.size(); i++) { - v3.push_back(v1[i] * v2[i]); + template + inline void operator()(Dim& self) const { + UnrollMul::Run(d1_, d2_, self.data()); } - return make_ddim(v3); + const int64_t* d1_; + const int64_t* d2_; +}; + +DDim DDim::operator*(const DDim& d) const { + PADDLE_ENFORCE(rank_ == d.rank_); + DDim ret; + ret.rank_ = rank_; + ret.apply_visitor(DDimMulVisitor(data(), d.data())); + return ret; } int64_t get(const DDim& ddim, int idx) { return ddim[idx]; } -void set(DDim& ddim, int idx, int value) { ddim[idx] = value; } - -/// @cond HIDDEN -struct VectorizeVisitor : public boost::static_visitor<> { - std::vector& vector; - - explicit VectorizeVisitor(std::vector& v) : vector(v) {} - - template - void operator()(const T& t) { - vector.push_back(t.head); - this->operator()(t.tail); - } - - void operator()(const Dim<0>& t) {} -}; -/// @endcond +void set(DDim& ddim, int idx, int value) { ddim[idx] = value; } // NOLINT std::vector vectorize(const DDim& ddim) { - std::vector result; - VectorizeVisitor visitor(result); - boost::apply_visitor(visitor, ddim); + std::vector result(DDim::kMaxRank); + for (int i = 0; i < ddim.size(); ++i) { + result[i] = ddim[i]; + } + result.resize(ddim.size()); return result; } // NOTE: framework::vectorize converts to type int64_t // which does not fit cudnn inputs. std::vector vectorize2int(const DDim& ddim) { - std::vector temp = vectorize(ddim); - std::vector result(temp.begin(), temp.end()); + std::vector result(DDim::kMaxRank); + for (int i = 0; i < ddim.size(); ++i) { + result[i] = ddim[i]; + } + result.resize(ddim.size()); return result; } -struct ProductVisitor : public boost::static_visitor { +struct ProductVisitor { template int64_t operator()(const Dim& dim) { return product(dim); @@ -220,65 +150,27 @@ struct ProductVisitor : public boost::static_visitor { }; int64_t product(const DDim& ddim) { - ProductVisitor visitor; - return boost::apply_visitor(visitor, ddim); + return ddim.apply_visitor(ProductVisitor()); } -struct SliceVectorizeVisitor : public boost::static_visitor<> { - std::vector& vector; - int begin; - int end; - - SliceVectorizeVisitor(std::vector& v, int b, int e) - : vector(v), begin(b), end(e) { - PADDLE_ENFORCE(begin < end, - "Begin index must be less than end index in ddim slice."); - PADDLE_ENFORCE(begin >= 0, - "Begin index can't be less than zero in ddim slice."); - } - - template - void operator()(const Dim& dim) { - if (begin == 0) { - vector.push_back(dim.head); - } else { - --begin; - } - --end; - if (end > 0) { - this->operator()(dim.tail); - } - } - - void operator()(const Dim<0>& dim) { - PADDLE_ENFORCE(end == 0, "End index in ddim slice is out of bound."); - } -}; - DDim slice_ddim(const DDim& dim, int begin, int end) { - std::vector vec; - vec.reserve(end - begin); - SliceVectorizeVisitor visitor(vec, begin, end); - boost::apply_visitor(visitor, dim); - return make_ddim(vec); -} - -/// \cond HIDDEN - -struct ArityVisitor : boost::static_visitor { - template - int operator()(Dim) const { - return D; + PADDLE_ENFORCE(begin < end, + "Begin index must be less than end index in ddim slice."); + PADDLE_ENFORCE(begin >= 0, + "Begin index can't be less than zero in ddim slice."); + DDim ret; + ret.rank_ = end - begin; + for (int i = 0; i < ret.rank_; ++i) { + ret[i] = dim[i + begin]; } -}; - -/// \endcond + return ret; +} -int arity(const DDim& d) { return boost::apply_visitor(ArityVisitor(), d); } +int arity(const DDim& d) { return d.size(); } /// \cond HIDDEN -struct DDimPrinter : boost::static_visitor { +struct DDimPrinter { std::ostream& os; explicit DDimPrinter(std::ostream& os_) : os(os_) {} @@ -291,15 +183,10 @@ struct DDimPrinter : boost::static_visitor { /// \endcond std::ostream& operator<<(std::ostream& os, const DDim& ddim) { - DDimPrinter printer(os); - boost::apply_visitor(printer, ddim); + ddim.apply_visitor(DDimPrinter(os)); return os; } -DDim::DDim(std::initializer_list init_list) { - *this = make_ddim(init_list); -} - DDim flatten_to_2d(const DDim& src, int num_col_dims) { int rank = src.size(); return make_ddim({product(slice_ddim(src, 0, num_col_dims)), @@ -309,21 +196,23 @@ DDim flatten_to_2d(const DDim& src, int num_col_dims) { DDim flatten_to_1d(const DDim& src) { return make_ddim({product(src)}); } DDim stride(const DDim& ddim) { - std::vector strides(ddim.size()); + DDim strides; + strides.rank_ = ddim.size(); strides[ddim.size() - 1] = 1; for (int i = ddim.size() - 2; i >= 0; --i) { strides[i] = strides[i + 1] * ddim[i + 1]; } - return framework::make_ddim(strides); + return strides; } DDim stride_numel(const framework::DDim& ddim) { - std::vector strides(ddim.size()); + DDim strides; + strides.rank_ = ddim.size(); strides[ddim.size() - 1] = ddim[ddim.size() - 1]; for (int i = ddim.size() - 2; i >= 0; --i) { strides[i] = strides[i + 1] * ddim[i]; } - return framework::make_ddim(strides); + return strides; } } // namespace framework diff --git a/paddle/fluid/framework/ddim.h b/paddle/fluid/framework/ddim.h index f05b5ee3fae..bff710040eb 100644 --- a/paddle/fluid/framework/ddim.h +++ b/paddle/fluid/framework/ddim.h @@ -18,8 +18,6 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/dim.h" -#include "paddle/fluid/platform/enforce.h" -#include "paddle/fluid/platform/variant.h" namespace paddle { namespace framework { @@ -29,51 +27,138 @@ namespace framework { * * The number of dimensions must be between [1, 9]. */ -struct DDim { - typedef boost::variant, Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, - Dim<7>, Dim<8>, Dim<9>> - DDimVar; - DDimVar var; +class DDim { + public: + constexpr static int kMaxRank = 9; - DDim() : var(Dim<1>()) {} + DDim() : rank_(1) { dim_[0] = 0; } + + DDim(const int* d, int n); + DDim(const int64_t* d, int n); template - explicit DDim(const Dim& in) : var(in) {} + /*implicit*/ DDim(const Dim& in) : rank_(D) { // NOLINT + UnsafeCast() = in; + } - /*implicit*/ DDim(std::initializer_list init_list); + /*implicit*/ DDim(std::initializer_list init_list) + : DDim(init_list.begin(), init_list.size()) {} template - DDim& operator=(const Dim& in) { - var = in; + inline DDim& operator=(const Dim& in) { + rank_ = D; + UnsafeCast() = in; return *this; } - int64_t& operator[](int idx); - int64_t operator[](int idx) const; + inline int64_t& operator[](int idx) { return dim_[idx]; } - template - typename Visitor::result_type apply_visitor(Visitor& visitor) { - return var.apply_visitor(visitor); + inline int64_t operator[](int idx) const { return dim_[idx]; } + + inline int64_t& at(int idx) { + PADDLE_ENFORCE(idx >= 0 && idx < rank_); + return dim_[idx]; } - template - typename Visitor::result_type apply_visitor(Visitor& visitor) const { - return var.apply_visitor(visitor); + inline int64_t at(int idx) const { + PADDLE_ENFORCE(idx >= 0 && idx < rank_); + return dim_[idx]; } - DDimVar getVar() { return var; } + template + typename std::result_of&)>::type apply_visitor( + Visitor&& visitor); + + template + typename std::result_of&)>::type apply_visitor( + Visitor&& visitor) const; + + bool operator==(const DDim& d) const; + + bool operator!=(const DDim& d) const; + + DDim operator+(const DDim& d) const; - bool operator==(DDim d) const; + DDim operator*(const DDim& d) const; - bool operator!=(DDim d) const; + // Make DDim act like std::vector + using iterator = int64_t*; + using const_iterator = const int64_t*; - DDim operator+(DDim d) const; + int64_t* data() { return dim_.data(); } + const int64_t* data() const { return dim_.data(); } - DDim operator*(DDim d) const; + iterator begin() { return data(); } + const_iterator begin() const { return data(); } + iterator end() { return data() + rank_; } + const_iterator end() const { return data() + rank_; } + + int size() const { return rank_; } + + private: + template + inline Dim& UnsafeCast() { + return const_cast&>(const_cast(this)->UnsafeCast()); + } - int size() const; + template + inline const Dim& UnsafeCast() const { + static_assert(M >= 0 && M <= kMaxRank, "Invalid rank"); + auto* p = static_cast(&dim_); + return *reinterpret_cast*>(p); + } + + friend DDim slice_ddim(const DDim& dim, int begin, int end); + friend DDim stride(const DDim& ddim); + friend DDim stride_numel(const DDim& ddim); + + Dim dim_; + int rank_; }; +#define PADDLE_VISIT_DDIM(rank) \ + case rank: \ + return visitor(UnsafeCast()) + +template +typename std::result_of&)>::type DDim::apply_visitor( + Visitor&& visitor) { + switch (rank_) { + PADDLE_VISIT_DDIM(0); + PADDLE_VISIT_DDIM(1); + PADDLE_VISIT_DDIM(2); + PADDLE_VISIT_DDIM(3); + PADDLE_VISIT_DDIM(4); + PADDLE_VISIT_DDIM(5); + PADDLE_VISIT_DDIM(6); + PADDLE_VISIT_DDIM(7); + PADDLE_VISIT_DDIM(8); + PADDLE_VISIT_DDIM(9); + default: + PADDLE_THROW("Invalid rank %d", rank_); + } +} + +template +typename std::result_of&)>::type DDim::apply_visitor( + Visitor&& visitor) const { + switch (rank_) { + PADDLE_VISIT_DDIM(0); + PADDLE_VISIT_DDIM(1); + PADDLE_VISIT_DDIM(2); + PADDLE_VISIT_DDIM(3); + PADDLE_VISIT_DDIM(4); + PADDLE_VISIT_DDIM(5); + PADDLE_VISIT_DDIM(6); + PADDLE_VISIT_DDIM(7); + PADDLE_VISIT_DDIM(8); + PADDLE_VISIT_DDIM(9); + default: + PADDLE_THROW("Invalid rank %d", rank_); + } +} +#undef PADDLE_VISIT_DDIM + /** * \brief Make a DDim from std::vector * @@ -92,7 +177,7 @@ DDim make_ddim(const std::vector& dims); DDim make_ddim(std::initializer_list dims); int64_t get(const DDim& dim, int idx); -void set(DDim& dim, int idx, int val); +void set(DDim& dim, int idx, int val); // NOLINT std::vector vectorize(const DDim& ddim); std::vector vectorize2int(const DDim& ddim); @@ -129,12 +214,3 @@ DDim stride(const DDim& ddim); DDim stride_numel(const DDim& ddim); } // namespace framework } // namespace paddle - -namespace boost { - -template -T get(const paddle::framework::DDim& in) { - return boost::get(in.var); -} - -} // namespace boost diff --git a/paddle/fluid/framework/dim.h b/paddle/fluid/framework/dim.h index 73f92fa389f..3ae60a3119e 100644 --- a/paddle/fluid/framework/dim.h +++ b/paddle/fluid/framework/dim.h @@ -16,328 +16,184 @@ #include #include #include +#include #include +#include "paddle/fluid/framework/array.h" #include "paddle/fluid/platform/assert.h" +#include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/hostdevice.h" namespace paddle { namespace framework { // Statically sized, statically indexed dimension -template -struct Dim { - static constexpr int dimensions = i; +template +class Dim : public Array { + public: + static_assert(N >= 0, "N must be not less than 0"); - template - HOSTDEVICE Dim(int64_t _head, Args... _tail) : head(_head), tail(_tail...) { - static_assert(sizeof...(_tail) == i - 1, - "Dim initialized with the wrong number of parameters"); - } + static constexpr int kRank = N; + using BaseClass = Array; - HOSTDEVICE - Dim(int64_t _head, const Dim& _tail) : head(_head), tail(_tail) {} + inline Dim(int64_t head, const Dim& tail) { + (*this)[0] = head; + new (this->GetMutable() + 1) Dim(tail); + } - HOSTDEVICE - Dim() : head(0), tail() {} + template + HOSTDEVICE explicit Dim(int64_t head, Args... args) + : BaseClass(head, args...) {} /** Construct a Dim from a linear index and size. Uses Fortran order * indexing. */ - HOSTDEVICE - Dim(int64_t idx, const Dim& size) - : head(idx % size.head), tail(idx / size.head, size.tail) {} + HOSTDEVICE Dim(int64_t idx, const Dim& size); /** Construct a Dim with each dimension set to the given index */ - HOSTDEVICE - Dim(int64_t idx) : head(idx), tail(idx) {} + HOSTDEVICE explicit Dim(int64_t idx) { this->Fill(idx); } - HOSTDEVICE - bool operator==(const Dim& o) const { - return (head == o.head) && (tail == o.tail); - } + HOSTDEVICE Dim() = default; - HOSTDEVICE - bool operator!=(const Dim& o) const { return !(*this == o); } + HOSTDEVICE int64_t* data() { return this->GetMutable(); } - HOSTDEVICE - int64_t& operator[](int idx); - HOSTDEVICE - int64_t operator[](int idx) const; + HOSTDEVICE const int64_t* data() const { return this->Get(); } HOST std::string to_string() const; - - int64_t head; - Dim tail; -}; - -// Base case specialization -template <> -struct Dim<0> { - static constexpr int dimensions = 0; - - HOSTDEVICE - Dim(int64_t _head) {} - - HOSTDEVICE - Dim() {} - - HOSTDEVICE - Dim(int idx, const Dim<0>& size) { -#ifndef __CUDA_ARCH__ - if (idx > 0) { - throw std::invalid_argument("Index out of range."); - } -#else - PADDLE_ASSERT(idx == 0); -#endif - } - - HOSTDEVICE - bool operator==(const Dim<0>& o) const { return true; } - - HOSTDEVICE - bool operator!=(const Dim<0>& o) const { return false; } - - HOSTDEVICE - int64_t& operator[](int idx); - HOSTDEVICE - int64_t operator[](int idx) const; }; -namespace { - -// Helper for accessing Dim classes -template -struct DimGetter { - // Return a copy if Dim is const - template - HOSTDEVICE static int64_t impl(const D& d) { - return DimGetter::impl(d.tail); - } - // Return a reference if Dim is mutable - template - HOSTDEVICE static int64_t& impl(D& d) { - return DimGetter::impl(d.tail); +namespace detail { +template +struct FortranOrderIndexingConstructorFunctor { + HOSTDEVICE inline static void Run(const int64_t* in, int64_t* idx, + int64_t* out) { + out[kStart] = (*idx) % in[kStart]; + (*idx) /= in[kStart]; + FortranOrderIndexingConstructorFunctor::Run(in, idx, + out); } }; -// Eureka! We found the element! -template <> -struct DimGetter<0> { - // Return a copy if Dim is const - template - HOSTDEVICE static int64_t impl(const D& d) { - return d.head; - } - // Return a reference if Dim is mutable - template - HOSTDEVICE static int64_t& impl(D& d) { - return d.head; - } +template +struct FortranOrderIndexingConstructorFunctor { + HOSTDEVICE inline static void Run(const int64_t* in, int64_t* idx, + int64_t* out) {} }; +} // namespace detail -template -HOSTDEVICE int64_t& indexer(Dim& dim, int idx) { -#ifndef __CUDA_ARCH__ - if (idx < 0) { - throw std::invalid_argument("Tried to access a negative dimension"); - } -#else - PADDLE_ASSERT(idx >= 0); -#endif - if (idx == 0) { - return dim.head; - } - return indexer(dim.tail, idx - 1); +template +HOSTDEVICE Dim::Dim(int64_t idx, const Dim& size) { + detail::FortranOrderIndexingConstructorFunctor<0, N, N == 0>::Run( + size.Get(), &idx, this->GetMutable()); } -template <> -HOSTDEVICE int64_t& indexer<0>(Dim<0>& dim, int idx) { -#ifndef __CUDA_ARCH__ - throw std::invalid_argument("Invalid index"); -#else - PADDLE_ASSERT(false); -#if CUDA_VERSION < 8000 - // On CUDA versions previous to 8.0, only __shared__ variables - // could be declared as static in the device code. - int64_t head = 0; -#else - static int64_t head = 0; -#endif - return head; -#endif +template +HOSTDEVICE inline int64_t get(const Dim& dim) { + return dim[idx]; } -template -HOSTDEVICE int64_t indexer(const Dim& dim, int idx) { -#ifndef __CUDA_ARCH__ - if (idx < 0) { - throw std::invalid_argument("Tried to access a negative dimension"); - } -#else - PADDLE_ASSERT(idx >= 0); -#endif - if (idx == 0) { - return dim.head; - } - return indexer(dim.tail, idx - 1); -} - -template <> -HOSTDEVICE int64_t indexer<0>(const Dim<0>& dim, int idx) { -#ifndef __CUDA_ARCH__ - throw std::invalid_argument("Invalid index"); -#else - PADDLE_ASSERT(false); -#if CUDA_VERSION < 8000 - // On CUDA versions previous to 8.0, only __shared__ variables - // could be declared as static in the device code. - int64_t head = 0; -#else - static int64_t head = 0; -#endif - return head; -#endif -} - -} // namespace -// Static access to constant Dim -template -HOSTDEVICE int64_t get(const Dim& d) { - return DimGetter::impl(d); -} - -// Static access to mutable Dim -template -HOSTDEVICE int64_t& get(Dim& d) { - return DimGetter::impl(d); -} - -// Dynamic access to constant Dim -template -HOSTDEVICE int64_t Dim::operator[](int i) const { - return indexer(*this, i); +template +HOSTDEVICE inline int64_t& get(Dim& dim) { // NOLINT + return dim[idx]; } -// Dynamic access to mutable Dim -template -HOSTDEVICE int64_t& Dim::operator[](int i) { - return indexer(*this, i); +template +HOSTDEVICE inline int64_t get(const Dim& dim, int idx) { + return dim[idx]; } -// Dynamic access to constant Dim -inline HOSTDEVICE int64_t Dim<0>::operator[](int i) const { - return indexer(*this, i); -} - -// Dynamic access to mutable Dim -inline HOSTDEVICE int64_t& Dim<0>::operator[](int i) { - return indexer(*this, i); -} - -// Dynamic access to constant Dim -// without std::enable_if will try to instantiate this on get<0>(d) -template -HOSTDEVICE typename std::enable_if<(l > 0), int64_t>::type get(const Dim& d, - int i) { - return d[i]; -} - -// Dynamic access to mutable Dim -template -HOSTDEVICE typename std::enable_if<(l > 0), int64_t&>::type get(Dim& d, - int i) { - return d[i]; +template +HOSTDEVICE inline int64_t& get(Dim& dim, int idx) { // NOLINT + return dim[idx]; } // Dot product of two dims -template -HOSTDEVICE int64_t linearize(const Dim& a, const Dim& b) { - return a.head * b.head + linearize(a.tail, b.tail); -} - -// Base case dot product of two Dims -// Notice it is inline because it is no longer a template -template <> -HOSTDEVICE inline int64_t linearize(const Dim<0>& a, const Dim<0>& b) { - return 0; +template +HOSTDEVICE inline int64_t linearize(const Dim& a, const Dim& b) { + return UnrollProduct::Run(a.Get(), b.Get()); } // Product of a Dim -template -HOSTDEVICE int64_t product(const Dim& a, int prod = 1) { - return prod * a.head * product(a.tail); -} - -// Base case product of a Dim -// Notice it is inline because it is no longer a template -template <> -HOSTDEVICE inline int64_t product(const Dim<0>& a, int prod) { - return prod; +template +HOSTDEVICE inline int64_t product(const Dim& a) { + return UnrollProduct::Run(a.Get()); } // Is 0 <= idx_i < size_i for all i? -template -HOSTDEVICE bool contained(const Dim& idx, const Dim& size) { - return ((0 <= idx.head) && (idx.head < size.head) && - contained(idx.tail, size.tail)); -} +namespace detail { +template +struct ContainedFunctor { + HOSTDEVICE static inline bool Run(const int64_t* idx, const int64_t* size) { + return (idx[kStart] >= 0 && idx[kStart] < size[kStart]) && + ContainedFunctor::Run(idx, + size); + } +}; + +template +struct ContainedFunctor { + HOSTDEVICE static constexpr inline bool Run(const int64_t* idx, + const int64_t* size) { + return true; + } +}; +} // namespace detail -// Base case of is 0 <= idx_i < size_i ? -// Notice it is inline because it is no longer a template -template <> -HOSTDEVICE inline bool contained(const Dim<0>& idx, const Dim<0>& size) { - return true; +template +HOSTDEVICE inline bool contained(const Dim& idx, const Dim& size) { + return detail::ContainedFunctor<0, N, N == 0>::Run(idx.Get(), size.Get()); } /** * \brief Compute exclusive prefix-multiply of a Dim. */ -template -HOSTDEVICE Dim ex_prefix_mul(const Dim& src, int mul = 1) { - return Dim(mul, ex_prefix_mul(src.tail, mul * src.head)); -} +namespace detail { +template +struct ExPrefixMulFunctor { + HOSTDEVICE static inline void Run(const int64_t* in, int64_t* out) { + kStart == 0 ? out[kStart] = 1 : out[kStart] = + out[kStart - 1] * in[kStart - 1]; + detail::ExPrefixMulFunctor::Run(in, + out); + } +}; + +template +struct ExPrefixMulFunctor { + HOSTDEVICE static inline void Run(const int64_t* in, int64_t* out) {} +}; +} // namespace detail -///\cond HIDDEN -// Base case of ex_prefix_mul -// Notice it is inline because it is no longer a template -template <> -HOSTDEVICE inline Dim<0> ex_prefix_mul(const Dim<0>& src, int mul) { - return Dim<0>(); +template +HOSTDEVICE inline Dim ex_prefix_mul(const Dim& src) { + Dim ret; + detail::ExPrefixMulFunctor<0, N, N == 0>::Run(src.Get(), ret.GetMutable()); + return ret; } -///\endcond /** * Add two dimensions together */ -template -HOSTDEVICE Dim dim_plus(const Dim& a, const Dim& b) { - return Dim(a.head + b.head, dim_plus(a.tail, b.tail)); +template +HOSTDEVICE inline Dim dim_plus(const Dim& a, const Dim& b) { + Dim ret; + UnrollAdd::Run(a.Get(), b.Get(), ret.GetMutable()); + return ret; } -// Base case -template <> -HOSTDEVICE inline Dim<0> dim_plus(const Dim<0>& a, const Dim<0>& b) { - return Dim<0>(); -} - -template -HOSTDEVICE Dim operator+(const Dim& lhs, const Dim& rhs) { +template +HOSTDEVICE inline Dim operator+(const Dim& lhs, const Dim& rhs) { return dim_plus(lhs, rhs); } /** * Multiply two dimensions together */ -template -HOSTDEVICE Dim dim_mult(const Dim& a, const Dim& b) { - return Dim(a.head * b.head, dim_mult(a.tail, b.tail)); -} - -// Base case -template <> -HOSTDEVICE inline Dim<0> dim_mult(const Dim<0>& a, const Dim<0>& b) { - return Dim<0>(); +template +HOSTDEVICE inline Dim dim_mult(const Dim& a, const Dim& b) { + Dim ret; + UnrollMul::Run(a.Get(), b.Get(), ret.GetMutable()); + return ret; } template @@ -354,23 +210,32 @@ HOSTDEVICE Dim operator*(const Dim& lhs, const Dim& rhs) { * \return Dim object the same size as \p size with normalized strides * */ +namespace detail { +template +struct NormalizeStridesFunctor { + HOSTDEVICE static void Run(const int64_t* size, const int64_t* stride, + int64_t* ret) { + ret[kStart] = (size[kStart] == 1 ? 0 : stride[kStart]); + NormalizeStridesFunctor::Run( + size, stride, ret); + } +}; -template -HOSTDEVICE Dim normalize_strides(const Dim& size, const Dim& stride) { - int norm_stride = size.head == 1 ? 0 : stride.head; - return Dim(norm_stride, normalize_strides(size.tail, stride.tail)); -} - -///\cond HIDDEN +template +struct NormalizeStridesFunctor { + HOSTDEVICE static void Run(const int64_t* size, const int64_t* stride, + int64_t* ret) {} +}; +} // namespace detail -template <> -HOSTDEVICE inline Dim<0> normalize_strides(const Dim<0>& size, - const Dim<0>& stride) { - return Dim<0>(); +template +HOSTDEVICE Dim normalize_strides(const Dim& size, const Dim& stride) { + Dim ret; + detail::NormalizeStridesFunctor<0, N, N == 0>::Run(size.Get(), stride.Get(), + ret.GetMutable()); + return ret; } -///\endcond - /** * Helper function to create a Dim * @@ -379,25 +244,17 @@ HOSTDEVICE inline Dim<0> normalize_strides(const Dim<0>& size, */ template -HOSTDEVICE Dim make_dim(Args... idxes) { +HOSTDEVICE inline Dim make_dim(Args... idxes) { return Dim(idxes...); } // Allows us to output a Dim -// XXX For some reason, overloading fails to resolve this correctly -template -typename std::enable_if<(i > 1), std::ostream&>::type operator<<( - std::ostream& os, const Dim& d) { - os << d.head << ", " << d.tail; - return os; -} - -// Base case that allows us to output a Dim -// XXX I wish this could be an overload instead of a template -template -typename std::enable_if<(i == 1), std::ostream&>::type operator<<( - std::ostream& os, const Dim& d) { - os << d.head; +template +inline std::ostream& operator<<(std::ostream& os, const Dim& d) { + os << d[0]; + for (int i = 1; i < N; ++i) { + os << ", " << d[i]; + } return os; } @@ -405,25 +262,23 @@ inline std::ostream& operator<<(std::ostream& os, const Dim<0>& d) { return os; } -template -HOST std::string Dim::to_string() const { +template +HOST std::string Dim::to_string() const { std::stringstream stream; - stream << *this; - return stream.str(); } -template -HOSTDEVICE Dim linear_to_dimension(int linear_index, Dim extents) { - Dim result; +template +HOSTDEVICE Dim linear_to_dimension(int linear_index, const Dim& extents) { + Dim result; - for (int i = 0; i < D - 1; ++i) { + for (int i = 0; i < N - 1; ++i) { result[i] = linear_index % extents[i]; linear_index /= extents[i]; } - result[D - 1] = linear_index; + result[N - 1] = linear_index; return result; } diff --git a/paddle/fluid/framework/dlpack_tensor.cc b/paddle/fluid/framework/dlpack_tensor.cc index 04e3f78afe4..5014fcd06a0 100644 --- a/paddle/fluid/framework/dlpack_tensor.cc +++ b/paddle/fluid/framework/dlpack_tensor.cc @@ -62,7 +62,7 @@ static DLDataType GetDLDataTypeFromTypeIndex(const std::type_index &type) { struct DLContextVisitor : public boost::static_visitor<::DLContext> { inline ::DLContext operator()(const platform::CPUPlace &place) const { - DLContext ctx; + ::DLContext ctx; ctx.device_type = kDLCPU; ctx.device_id = 0; return ctx; @@ -70,7 +70,7 @@ struct DLContextVisitor : public boost::static_visitor<::DLContext> { inline ::DLContext operator()(const platform::CUDAPlace &place) const { #ifdef PADDLE_WITH_CUDA - DLContext ctx; + ::DLContext ctx; ctx.device_type = kDLGPU; ctx.device_id = place.device; return ctx; @@ -81,7 +81,7 @@ struct DLContextVisitor : public boost::static_visitor<::DLContext> { inline ::DLContext operator()(const platform::CUDAPinnedPlace &place) const { #ifdef PADDLE_WITH_CUDA - DLContext ctx; + ::DLContext ctx; ctx.device_type = kDLCPUPinned; ctx.device_id = 0; return ctx; diff --git a/paddle/fluid/framework/dlpack_tensor.h b/paddle/fluid/framework/dlpack_tensor.h index 0c52bce1ef6..e48b0d5c88f 100644 --- a/paddle/fluid/framework/dlpack_tensor.h +++ b/paddle/fluid/framework/dlpack_tensor.h @@ -38,7 +38,7 @@ class DLPackTensor { // The shape in DLTensor is defined as int64_t* // Add this member to make TVMTensor init without heap allocation - ShapeType shape_[9]; + ShapeType shape_[DDim::kMaxRank]; }; } // namespace framework diff --git a/paddle/fluid/framework/unroll_array_ops.h b/paddle/fluid/framework/unroll_array_ops.h new file mode 100644 index 00000000000..fb0a89530f6 --- /dev/null +++ b/paddle/fluid/framework/unroll_array_ops.h @@ -0,0 +1,169 @@ +// 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. + +#pragma once +#include +#include "paddle/fluid/platform/hostdevice.h" + +namespace paddle { +namespace framework { + +namespace detail { + +template +struct UnrollFillConstant { + template + HOSTDEVICE inline static void Run(T *data, T val) { + data[kStart] = val; + UnrollFillConstant::Run(data, val); + } +}; + +template +struct UnrollFillConstant { + template + HOSTDEVICE inline static void Run(T *data, T val) {} +}; + +template +struct UnrollAssign { + template + HOSTDEVICE inline static void Run(const Tin *d1, Tout *d2) { + d2[kStart] = static_cast(d1[kStart]); + UnrollAssign::Run(d1, d2); + } +}; + +template +struct UnrollAssign { + template + HOSTDEVICE inline static void Run(const Tin *d1, Tout *d2) {} +}; + +template +struct UnrollVarArgsAssign { + template + HOSTDEVICE inline static void Run(T *d, T val, Args... args) { + static_assert(sizeof...(args) + 1 == kEnd - kStart, "Wrong argument"); + d[kStart] = val; + UnrollVarArgsAssign::Run(d, + args...); + } +}; + +template +struct UnrollVarArgsAssign { + HOSTDEVICE inline static void Run(T *d) {} +}; + +template +struct UnrollCompare { + template + HOSTDEVICE inline static bool Run(const T *d1, const T *d2) { + return d1[kStart] == d2[kStart] && + UnrollCompare::Run(d1, d2); + } +}; + +template +struct UnrollCompare { + template + HOSTDEVICE inline constexpr static bool Run(const T *d1, const T *d2) { + return true; + } +}; + +template +struct UnrollAdd { + template + HOSTDEVICE inline static void Run(const T *d1, const T *d2, T *d3) { + d3[kStart] = d1[kStart] + d2[kStart]; + UnrollAdd::Run(d1, d2, d3); + } +}; + +template +struct UnrollAdd { + template + HOSTDEVICE inline static void Run(const T *d1, const T *d2, T *d3) {} +}; + +template +struct UnrollMul { + template + HOSTDEVICE inline static void Run(const T *d1, const T *d2, T *d3) { + d3[kStart] = d1[kStart] * d2[kStart]; + UnrollMul::Run(d1, d2, d3); + } +}; + +template +struct UnrollMul { + template + HOSTDEVICE inline static void Run(const T *d1, const T *d2, T *d3) {} +}; + +template +struct UnrollProduct { + template + HOSTDEVICE inline static T Run(const T *d) { + return d[kStart] * + UnrollProduct::Run(d); + } + + template + HOSTDEVICE inline static T Run(const T *d1, const T *d2) { + return d1[kStart] * d2[kStart] + + UnrollProduct::Run(d1, d2); + } +}; + +template +struct UnrollProduct { + template + HOSTDEVICE inline constexpr static T Run(const T *d) { + return 1; + } + + template + HOSTDEVICE inline constexpr static T Run(const T *d1, const T *d2) { + return 0; + } +}; + +} // namespace detail + +template +using UnrollFillConstant = detail::UnrollFillConstant<0, N, N == 0>; + +template +using UnrollAssign = detail::UnrollAssign<0, N, N == 0>; + +template +using UnrollVarArgsAssign = detail::UnrollVarArgsAssign; + +template +using UnrollCompare = detail::UnrollCompare<0, N, N == 0>; + +template +using UnrollAdd = detail::UnrollAdd<0, N, N == 0>; + +template +using UnrollMul = detail::UnrollMul<0, N, N == 0>; + +template +using UnrollProduct = detail::UnrollProduct<0, N, N == 0>; + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/operators/controlflow/logical_op.cc b/paddle/fluid/operators/controlflow/logical_op.cc index 6446cab5ec5..2e7f3edd55c 100644 --- a/paddle/fluid/operators/controlflow/logical_op.cc +++ b/paddle/fluid/operators/controlflow/logical_op.cc @@ -86,8 +86,6 @@ class UnaryLogicalOpInferShape : public framework::InferShapeBase { OpComment comment; PADDLE_ENFORCE(context->HasInput("X"), "Input(X) of %s operator must not be null", comment.type); - auto dim_x = context->GetInputDim("X"); - context->SetOutputDim("Out", context->GetInputDim("X")); context->ShareLoD("X", "Out"); } diff --git a/paddle/fluid/operators/crop_op.h b/paddle/fluid/operators/crop_op.h index 2d7d33bd4f9..cfc2cac7beb 100644 --- a/paddle/fluid/operators/crop_op.h +++ b/paddle/fluid/operators/crop_op.h @@ -68,7 +68,6 @@ void CropFunction(const framework::ExecutionContext& context) { } out->mutable_data(out_dims, context.GetPlace()); auto x_stride = framework::stride(x->dims()); - auto out_stride = framework::stride(out->dims()); auto offsets = GetOffsets(context); int64_t offset = 0; for (size_t i = 0; i < offsets.size(); ++i) { diff --git a/paddle/fluid/operators/cudnn_lstm_op.cu.cc b/paddle/fluid/operators/cudnn_lstm_op.cu.cc index dd64cc327fc..744d149714c 100644 --- a/paddle/fluid/operators/cudnn_lstm_op.cu.cc +++ b/paddle/fluid/operators/cudnn_lstm_op.cu.cc @@ -378,7 +378,6 @@ class CudnnLSTMGPUGradKernel : public framework::OpKernel { ->GetMutable(); auto input_dims = input->dims(); - auto weight_dims = weight->dims(); auto init_h_dims = init_h->dims(); auto init_c_dims = init_c->dims(); in_grad->mutable_data(ctx.GetPlace()); diff --git a/paddle/fluid/operators/detail/strided_memcpy.h b/paddle/fluid/operators/detail/strided_memcpy.h index 0b7c470fe72..fc223ce5593 100644 --- a/paddle/fluid/operators/detail/strided_memcpy.h +++ b/paddle/fluid/operators/detail/strided_memcpy.h @@ -27,8 +27,8 @@ struct StridedMemcpyFunctor; template struct StridedMemcpyFunctor { void operator()(const platform::DeviceContext& dev_ctx, const T* src, - framework::Dim<0> src_stride, framework::Dim<0> dst_dim, - framework::Dim<0> dst_stride, T* dst) const { + const int64_t* src_stride, const int64_t* dst_dim, + const int64_t* dst_stride, T* dst) const { auto place = dev_ctx.GetPlace(); if (platform::is_cpu_place(place)) { auto& cpu_place = boost::get(place); @@ -50,18 +50,18 @@ struct StridedMemcpyFunctor { template struct StridedMemcpyFunctor { void operator()(const platform::DeviceContext& dev_ctx, const T* src, - framework::Dim<1> src_stride, framework::Dim<1> dst_dim, - framework::Dim<1> dst_stride, T* dst) const { + const int64_t* src_stride, const int64_t* dst_dim, + const int64_t* dst_stride, T* dst) const { auto place = dev_ctx.GetPlace(); if (platform::is_cpu_place(place)) { auto& cpu_place = boost::get(place); - memory::Copy(cpu_place, dst, cpu_place, src, sizeof(T) * dst_dim.head); + memory::Copy(cpu_place, dst, cpu_place, src, sizeof(T) * dst_dim[0]); } else { #ifdef PADDLE_WITH_CUDA auto& gpu_place = boost::get(place); auto& cuda_ctx = reinterpret_cast(dev_ctx); - memory::Copy(gpu_place, dst, gpu_place, src, sizeof(T) * dst_dim.head, + memory::Copy(gpu_place, dst, gpu_place, src, sizeof(T) * dst_dim[0], cuda_ctx.stream()); #else PADDLE_THROW("Paddle is not compiled with GPU"); @@ -73,19 +73,19 @@ struct StridedMemcpyFunctor { template struct StridedMemcpyFunctor { void operator()(const platform::DeviceContext& dev_ctx, const T* src, - framework::Dim src_stride, framework::Dim dst_dim, - framework::Dim dst_stride, T* dst) const { - for (int64_t i = 0; i < dst_dim.head; ++i) { + const int64_t* src_stride, const int64_t* dst_dim, + const int64_t* dst_stride, T* dst) const { + for (int64_t i = 0; i < dst_dim[0]; ++i) { StridedMemcpyFunctor func; - func(dev_ctx, src, src_stride.tail, dst_dim.tail, dst_stride.tail, dst); - src += src_stride.head; - dst += dst_stride.head; + func(dev_ctx, src, src_stride + 1, dst_dim + 1, dst_stride + 1, dst); + src += src_stride[0]; + dst += dst_stride[0]; } } }; template -struct StridedCopyDimVisitor : public boost::static_visitor { +struct StridedCopyDimVisitor { StridedCopyDimVisitor(const platform::DeviceContext& dev_ctx, const T* src, const framework::DDim& src_stride, const framework::DDim& dst_stride, T* dst) @@ -95,13 +95,11 @@ struct StridedCopyDimVisitor : public boost::static_visitor { dst_stride_(dst_stride), dst_(dst) {} - template - void operator()(Dim dst_dim) const { - Dim src_stride = boost::get(src_stride_); - Dim dst_stride = boost::get(dst_stride_); - constexpr int dim = Dim::dimensions; - StridedMemcpyFunctor functor; - functor(dev_ctx_, src_, src_stride, dst_dim, dst_stride, dst_); + template + void operator()(const framework::Dim& dst_dim) const { + StridedMemcpyFunctor functor; + functor(dev_ctx_, src_, src_stride_.data(), dst_dim.data(), + dst_stride_.data(), dst_); } const platform::DeviceContext& dev_ctx_; diff --git a/paddle/fluid/operators/detection/generate_proposal_labels_op.cc b/paddle/fluid/operators/detection/generate_proposal_labels_op.cc index fddd6884017..a652d4d9575 100644 --- a/paddle/fluid/operators/detection/generate_proposal_labels_op.cc +++ b/paddle/fluid/operators/detection/generate_proposal_labels_op.cc @@ -64,8 +64,6 @@ class GenerateProposalLabelsOp : public framework::OperatorWithKernel { "Output(BboxOutsideWeights) of RpnTargetAssignOp should not be null"); auto rpn_rois_dims = ctx->GetInputDim("RpnRois"); - auto gt_classes_dims = ctx->GetInputDim("GtClasses"); - auto is_crowd_dims = ctx->GetInputDim("IsCrowd"); auto gt_boxes_dims = ctx->GetInputDim("GtBoxes"); auto im_info_dims = ctx->GetInputDim("ImInfo"); diff --git a/paddle/fluid/operators/detection/generate_proposals_op.cc b/paddle/fluid/operators/detection/generate_proposals_op.cc index 709c2dfc4b7..f1975a9a4be 100644 --- a/paddle/fluid/operators/detection/generate_proposals_op.cc +++ b/paddle/fluid/operators/detection/generate_proposals_op.cc @@ -53,12 +53,6 @@ class GenerateProposalsOp : public framework::OperatorWithKernel { PADDLE_ENFORCE(ctx->HasInput("Variances"), "Input(Variances) shouldn't be null."); - auto scores_dims = ctx->GetInputDim("Scores"); - auto bbox_deltas_dims = ctx->GetInputDim("BboxDeltas"); - auto im_info_dims = ctx->GetInputDim("ImInfo"); - auto anchors_dims = ctx->GetInputDim("Anchors"); - auto variances_dims = ctx->GetInputDim("Variances"); - ctx->SetOutputDim("RpnRois", {-1, 4}); ctx->SetOutputDim("RpnRoiProbs", {-1, 1}); } diff --git a/paddle/fluid/operators/detection/rpn_target_assign_op.cc b/paddle/fluid/operators/detection/rpn_target_assign_op.cc index 46fff9d338b..fd5d75ba527 100644 --- a/paddle/fluid/operators/detection/rpn_target_assign_op.cc +++ b/paddle/fluid/operators/detection/rpn_target_assign_op.cc @@ -58,7 +58,6 @@ class RpnTargetAssignOp : public framework::OperatorWithKernel { auto anchor_dims = ctx->GetInputDim("Anchor"); auto gt_boxes_dims = ctx->GetInputDim("GtBoxes"); - auto is_crowd_dims = ctx->GetInputDim("IsCrowd"); auto im_info_dims = ctx->GetInputDim("ImInfo"); PADDLE_ENFORCE_EQ(anchor_dims.size(), 2, "The rank of Input(Anchor) must be 2."); diff --git a/paddle/fluid/operators/elementwise/elementwise_op.h b/paddle/fluid/operators/elementwise/elementwise_op.h index 87bf7c6b156..775346c5524 100644 --- a/paddle/fluid/operators/elementwise/elementwise_op.h +++ b/paddle/fluid/operators/elementwise/elementwise_op.h @@ -178,7 +178,6 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { auto x_dims = ctx->GetInputDim("X"); auto y_dims = ctx->GetInputDim("Y"); - auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), "Rank of first input must >= rank of second input."); diff --git a/paddle/fluid/operators/expand_op.h b/paddle/fluid/operators/expand_op.h index 75dbf1d8bf5..33940824977 100644 --- a/paddle/fluid/operators/expand_op.h +++ b/paddle/fluid/operators/expand_op.h @@ -77,7 +77,6 @@ class ExpandKernel : public framework::OpKernel { auto& expand_times = context.Attr>("expand_times"); auto* out0 = context.Output("Out"); Eigen::DSizes bcast_dims; - auto x_dims = in0->dims(); for (size_t i = 0; i < expand_times.size(); ++i) { bcast_dims[i] = expand_times[i]; } diff --git a/paddle/fluid/operators/fc_op.cc b/paddle/fluid/operators/fc_op.cc index e80249fc878..7c53e5279da 100644 --- a/paddle/fluid/operators/fc_op.cc +++ b/paddle/fluid/operators/fc_op.cc @@ -148,7 +148,6 @@ class FCOpKernel : public framework::OpKernel { auto w = ctx.Input("W"); auto bias = ctx.Input("Bias"); auto output = ctx.Output("Out"); - auto in_dims = input->dims(); auto w_dims = w->dims(); auto out_dims = output->dims(); int M = framework::product(out_dims) / out_dims[out_dims.size() - 1]; diff --git a/paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc b/paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc index 1eb6523a2df..9344bfe65db 100644 --- a/paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc +++ b/paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc @@ -242,15 +242,15 @@ class FusedEmbeddingFCLSTMKernel : public framework::OpKernel { bool is_reverse = ctx.Attr("is_reverse"); \ bool use_peepholes = ctx.Attr("use_peepholes"); -#define INIT_BASE_SIZES \ - auto ids_dims = ids->dims(); /* T x M*/ \ - auto ids_numel = ids->numel(); /* T x 1*/ \ - auto wh_dims = wh->dims(); /* D x 4D*/ \ - const int D = wh_dims[0]; \ - const int D2 = D * 2; \ - const int D3 = D * 3; \ - int64_t row_number = embeddings->dims()[0]; \ - int64_t row_width = embeddings->dims()[1]; \ +#define INIT_BASE_SIZES \ + auto ids_dims = ids->dims(); /* T x M*/ \ + auto ids_numel = framework::product(ids_dims); /* T x 1*/ \ + auto wh_dims = wh->dims(); /* D x 4D*/ \ + const int D = wh_dims[0]; \ + const int D2 = D * 2; \ + const int D3 = D * 3; \ + int64_t row_number = embeddings->dims()[0]; \ + int64_t row_width = embeddings->dims()[1]; \ const int D4 = wh_dims[1]; #define INIT_BASE_INPUT_DATAS \ diff --git a/paddle/fluid/operators/hinge_loss_op.cc b/paddle/fluid/operators/hinge_loss_op.cc index 69e7fa4490b..f458ce6c83b 100644 --- a/paddle/fluid/operators/hinge_loss_op.cc +++ b/paddle/fluid/operators/hinge_loss_op.cc @@ -88,7 +88,6 @@ class HingeLossGradOp : public framework::OperatorWithKernel { "Input(Logits@GRAD) should not be null."); auto pred_dims = ctx->GetInputDim("Logits"); - auto lab_dims = ctx->GetInputDim("Labels"); auto loss_grad_dims = ctx->GetInputDim(framework::GradVarName("Loss")); PADDLE_ENFORCE_EQ(loss_grad_dims, pred_dims); diff --git a/paddle/fluid/operators/log_loss_op.cc b/paddle/fluid/operators/log_loss_op.cc index 9d248e03218..ef1fb83aa6e 100644 --- a/paddle/fluid/operators/log_loss_op.cc +++ b/paddle/fluid/operators/log_loss_op.cc @@ -92,7 +92,6 @@ class LogLossGradOp : public framework::OperatorWithKernel { "Output(Predicted@GRAD) should not be null."); auto pred_dims = ctx->GetInputDim("Predicted"); - auto label_dims = ctx->GetInputDim("Labels"); auto loss_grad_dims = ctx->GetInputDim(framework::GradVarName("Loss")); PADDLE_ENFORCE_EQ(loss_grad_dims, pred_dims); diff --git a/paddle/fluid/operators/math/math_function_impl.h b/paddle/fluid/operators/math/math_function_impl.h index 895a7019aa1..d1127ce4a24 100644 --- a/paddle/fluid/operators/math/math_function_impl.h +++ b/paddle/fluid/operators/math/math_function_impl.h @@ -37,9 +37,6 @@ void Transpose::operator()( for (int i = 0; i < Rank; i++) { permute[i] = axis[i]; } - auto in_dim = in.dims(); - auto out_dim = out->dims(); - auto eigen_in = framework::EigenTensor::From(in); auto eigen_out = framework::EigenTensor::From(*out); auto* dev = context.eigen_device(); diff --git a/paddle/fluid/operators/math/softmax_impl.h b/paddle/fluid/operators/math/softmax_impl.h index 9e99e44822b..1d9d98b1064 100644 --- a/paddle/fluid/operators/math/softmax_impl.h +++ b/paddle/fluid/operators/math/softmax_impl.h @@ -76,7 +76,6 @@ class SoftmaxFunctor> { void operator()(const DeviceContext& context, const framework::Tensor* X, framework::Tensor* Y) { auto in_dims = X->dims(); - auto out_dims = Y->dims(); const float* in_data = X->data(); float* out_data = Y->data(); const int kBatchDim = 0; diff --git a/paddle/fluid/operators/modified_huber_loss_op.cc b/paddle/fluid/operators/modified_huber_loss_op.cc index 35db4c1ad1f..9954e51083b 100644 --- a/paddle/fluid/operators/modified_huber_loss_op.cc +++ b/paddle/fluid/operators/modified_huber_loss_op.cc @@ -87,7 +87,6 @@ class ModifiedHuberLossGradOp : public framework::OperatorWithKernel { "Input(Out@Grad) must not be null."); auto x_dims = ctx->GetInputDim("X"); - auto y_dims = ctx->GetInputDim("Y"); auto intermediate_dims = ctx->GetInputDim("IntermediateVal"); auto out_grad_dims = ctx->GetInputDim(framework::GradVarName("Out")); diff --git a/paddle/fluid/operators/mul_op.cc b/paddle/fluid/operators/mul_op.cc index 8a111e6065b..154b5f0d08f 100644 --- a/paddle/fluid/operators/mul_op.cc +++ b/paddle/fluid/operators/mul_op.cc @@ -146,12 +146,6 @@ class MulGradOp : public framework::OperatorWithKernel { "Input(Out@GRAD) should not be null"); auto x_dims = ctx->GetInputDim("X"); auto y_dims = ctx->GetInputDim("Y"); - auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); - - auto x_mat_dims = framework::flatten_to_2d( - x_dims, ctx->Attrs().Get("x_num_col_dims")); - auto y_mat_dims = framework::flatten_to_2d( - y_dims, ctx->Attrs().Get("y_num_col_dims")); auto x_grad_name = framework::GradVarName("X"); auto y_grad_name = framework::GradVarName("Y"); diff --git a/paddle/fluid/operators/nce_op.cc b/paddle/fluid/operators/nce_op.cc index 9f97f7821dd..e58dccea131 100644 --- a/paddle/fluid/operators/nce_op.cc +++ b/paddle/fluid/operators/nce_op.cc @@ -36,7 +36,6 @@ class NCEOp : public framework::OperatorWithKernel { auto x_dims = ctx->GetInputDim("Input"); auto label_dims = ctx->GetInputDim("Label"); - auto w_dims = ctx->GetInputDim("Weight"); PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0]); int num_true_classes = label_dims.size() == 2 ? label_dims[1] : 1; if (ctx->HasInput("Bias")) { diff --git a/paddle/fluid/operators/norm_op.h b/paddle/fluid/operators/norm_op.h index d0224177ecf..6c95d3f3bf3 100644 --- a/paddle/fluid/operators/norm_op.h +++ b/paddle/fluid/operators/norm_op.h @@ -43,7 +43,6 @@ class NormKernel : public framework::OpKernel { out_norm->mutable_data(ctx.GetPlace()); auto xdim = in_x->dims(); - auto ndim = out_norm->dims(); T eps = static_cast(ctx.Attr("epsilon")); int axis = ctx.Attr("axis"); if (axis < 0) axis = xdim.size() + axis; diff --git a/paddle/fluid/operators/psroi_pool_op.h b/paddle/fluid/operators/psroi_pool_op.h index 1a424728f7f..5666613f6ef 100644 --- a/paddle/fluid/operators/psroi_pool_op.h +++ b/paddle/fluid/operators/psroi_pool_op.h @@ -41,7 +41,6 @@ class CPUPSROIPoolOpKernel : public framework::OpKernel { int rois_num = rois->dims()[0]; auto in_stride = framework::stride(in_dims); - auto roi_stride = framework::stride(rois->dims()); auto out_stride = framework::stride(out->dims()); const T* input_data = in->data(); diff --git a/paddle/fluid/operators/sequence_ops/sequence_slice_op.h b/paddle/fluid/operators/sequence_ops/sequence_slice_op.h index 03b59d71cc0..4bded0efb96 100644 --- a/paddle/fluid/operators/sequence_ops/sequence_slice_op.h +++ b/paddle/fluid/operators/sequence_ops/sequence_slice_op.h @@ -143,8 +143,6 @@ class SequenceSliceGradOpKernel : public framework::OpKernel { set_zero(ctx.template device_context(), x_grad, static_cast(0)); - auto out_grad_stride = framework::stride(out_grad->dims()); - for (size_t i = 0; i < out_lod[0].size() - 1; ++i) { Tensor out_grad_t = out_grad->Slice(static_cast(out_lod[0][i]), diff --git a/paddle/fluid/operators/strided_memcpy.h b/paddle/fluid/operators/strided_memcpy.h index c3d83a06f23..6a99ad9a90f 100644 --- a/paddle/fluid/operators/strided_memcpy.h +++ b/paddle/fluid/operators/strided_memcpy.h @@ -40,7 +40,7 @@ inline void StridedMemcpy(const platform::DeviceContext& dev_ctx, const T* src, const framework::DDim& dst_stride, T* dst) { paddle::operators::detail::StridedCopyDimVisitor func( dev_ctx, src, src_stride, dst_stride, dst); - boost::apply_visitor(func, dst_dim); + dst_dim.apply_visitor(func); } // Strided numel memory copy from src to dst by the specified axis -- GitLab