Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

【Sparse】add SparseTensor mv kernel(csr*dense_vec->dence_vec, coo*dense_vec->dense_vec) #43668

Merged
merged 2 commits into from Jun 28, 2022
Merged

【Sparse】add SparseTensor mv kernel(csr*dense_vec->dence_vec, coo*dense_vec->dense_vec) #43668

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c0e5a1b
[Sparse]add SparseTensor mv kernel(csr*dense_vec->dence_vec, coo*dens…
zhwesky2010 Jun 24, 2022
06dc589
fix CI
zhwesky2010 Jun 27, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
7 changes: 5 additions & 2 deletions paddle/fluid/platform/dynload/cusparse.h
View file
Open in desktop
Expand Up @@ -29,7 +29,6 @@ namespace dynload {
extern DynLoad__##__name __name

#if defined(PADDLE_WITH_CUDA)
// APIs available after CUDA 11.0
#if CUDA_VERSION >= 11000
#define CUSPARSE_ROUTINE_EACH(__macro) \
__macro(cusparseCreate); \
Expand All @@ -43,10 +42,14 @@ namespace dynload {
__macro(cusparseCreateCsr); \
__macro(cusparseCreateCoo); \
__macro(cusparseCreateDnMat); \
__macro(cusparseCreateDnVec); \
__macro(cusparseSpMM_bufferSize); \
__macro(cusparseSpMM); \
__macro(cusparseDestroySpMat); \
__macro(cusparseDestroyDnMat);
__macro(cusparseDestroyDnMat); \
__macro(cusparseDestroyDnVec); \
__macro(cusparseSpMV_bufferSize); \
__macro(cusparseSpMV);

CUSPARSE_ROUTINE_EACH(PLATFORM_DECLARE_DYNAMIC_LOAD_CUSPARSE_WRAP)
#endif
Expand Down
7 changes: 5 additions & 2 deletions paddle/phi/backends/dynload/cusparse.h
View file
Open in desktop
Expand Up @@ -41,7 +41,6 @@ extern void *cusparse_dso_handle;
extern DynLoad__##__name __name

#if defined(PADDLE_WITH_CUDA)
// APIs available after CUDA 11.0
#if CUDA_VERSION >= 11000
#define CUSPARSE_ROUTINE_EACH(__macro) \
__macro(cusparseCreate); \
Expand All @@ -55,10 +54,14 @@ extern void *cusparse_dso_handle;
__macro(cusparseCreateCsr); \
__macro(cusparseCreateCoo); \
__macro(cusparseCreateDnMat); \
__macro(cusparseCreateDnVec); \
__macro(cusparseSpMM_bufferSize); \
__macro(cusparseSpMM); \
__macro(cusparseDestroySpMat); \
__macro(cusparseDestroyDnMat);
__macro(cusparseDestroyDnMat); \
__macro(cusparseDestroyDnVec); \
__macro(cusparseSpMV_bufferSize); \
__macro(cusparseSpMV);

CUSPARSE_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUSPARSE_WRAP)
#endif
Expand Down
13 changes: 13 additions & 0 deletions paddle/phi/kernels/funcs/sparse/sparse_blas.h
View file
Open in desktop
Expand Up @@ -37,6 +37,14 @@ class SparseBlas {
T beta,
phi::DenseTensor* mat_out) const;

template <typename T, typename TensorType>
void SPMV(bool transa,
T alpha,
const TensorType& mat_a,
const phi::DenseTensor& vec_x,
T beta,
phi::DenseTensor* vec_out) const;

template <typename T, typename TensorType>
void SDDMM(bool transa,
bool transb,
Expand All @@ -60,6 +68,11 @@ class SparseBlasT : private SparseBlas<DeviceContext> {
Base()->template SPMM<T>(args...);
}

template <typename... ARGS>
void SPMV(ARGS... args) const {
Base()->template SPMV<T>(args...);
}

template <typename... ARGS>
void SDDMM(ARGS... args) const {
Base()->template SDDMM<T>(args...);
Expand Down
155 changes: 152 additions & 3 deletions paddle/phi/kernels/funcs/sparse/sparse_blas_impl.cu.h
View file
Open in desktop
Expand Up @@ -20,6 +20,7 @@
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/ddim.h"
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/core/enforce.h"
#include "paddle/phi/core/sparse_coo_tensor.h"
#include "paddle/phi/core/sparse_csr_tensor.h"
#include "paddle/phi/core/visit_type.h"
Expand Down Expand Up @@ -47,6 +48,8 @@ inline cusparseOperation_t GetTransposeOperation(const bool trans) {
}
}

/************* SPARSE MATRIX DESCRIPTOR (COO/CSR) ************/

template <typename T, typename IntT>
inline void CreateCsrDescriptor(const phi::SparseCsrTensor& x,
const phi::GPUContext& dev_ctx,
Expand Down Expand Up @@ -102,19 +105,83 @@ inline void CreateCsrDescriptor(const phi::SparseCsrTensor& x,
}
}

template <typename T, typename IntT>
inline void CreateCooDescriptor(const phi::SparseCooTensor& x,
const phi::GPUContext& dev_ctx,
cusparseSpMatDescr_t* descriptor) {
std::vector<int64_t> xdim_vec = phi::vectorize(x.dims());
auto x_ndims = xdim_vec.size();
PADDLE_ENFORCE_GE(
x_ndims,
2,
phi::errors::InvalidArgument("the dim size of SparseCsrTensor must be "
"greater than or eaqual to 2."));

int64_t M = xdim_vec[x_ndims - 2];
int64_t N = xdim_vec[x_ndims - 1];
int batch_size = 1;
for (int i = 0; i < x_ndims - 2; i++) {
batch_size *= xdim_vec[i];
}
int64_t nnz = x.nnz();

const IntT* indices_data = x.non_zero_indices().data<IntT>();
const T* values_data = x.non_zero_elements().data<T>();
auto rows_data = indices_data + (x_ndims - 2) * nnz;
auto cols_data = indices_data + (x_ndims - 1) * nnz;

int64_t batch_nnz = nnz / batch_size;
cudaDataType_t gpu_type = GetGpuDataType<T>();
dev_ctx.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseCreateCoo(descriptor,
M,
N,
batch_nnz,
const_cast<IntT*>(rows_data),
const_cast<IntT*>(cols_data),
const_cast<T*>(values_data),
CUSPARSE_INDEX_64I,
CUSPARSE_INDEX_BASE_ZERO,
gpu_type);
});

if (batch_size > 1) {
#if CUDA_VERSION >= 11070
dev_ctx.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseCooSetStridedBatch(
*descriptor, batch_size, batch_nnz);
});
#else
PADDLE_THROW(phi::errors::Unimplemented(
"Batch Sparse matmul use 'cusparseCooSetStridedBatch', which is "
"supported from CUDA 11.7"));
#endif
}
}

template <typename T>
class CuSparseSpMatDescriptor {
public:
explicit CuSparseSpMatDescriptor(const phi::SparseCsrTensor& x,
const phi::GPUContext& dev_ctx)
: dev_ctx_(dev_ctx) {
PD_VISIT_INTEGRAL_TYPES(
x.non_zero_crows().dtype(), "CuSparseSpMatDescriptor", ([&] {
x.non_zero_crows().dtype(), "Csr CuSparseSpMatDescriptor", ([&] {
CreateCsrDescriptor<T, data_t>(x, dev_ctx_, &descriptor_);
}));
VLOG(6) << "Create csr cusparseSpMatDescr_t " << &descriptor_;
}

explicit CuSparseSpMatDescriptor(const phi::SparseCooTensor& x,
const phi::GPUContext& dev_ctx)
: dev_ctx_(dev_ctx) {
PD_VISIT_INTEGRAL_TYPES(
x.non_zero_indices().dtype(), "Coo CuSparseSpMatDescriptor", ([&] {
CreateCooDescriptor<T, data_t>(x, dev_ctx_, &descriptor_);
}));
VLOG(6) << "Create coo cusparseSpMatDescr_t " << &descriptor_;
}

~CuSparseSpMatDescriptor() {
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseDestroySpMat(descriptor_);
Expand All @@ -129,6 +196,7 @@ class CuSparseSpMatDescriptor {
cusparseSpMatDescr_t descriptor_;
};

/************* DENSE MATRIX DESCRIPTOR ************/
template <typename T>
class CuSparseDnMatDescriptor {
public:
Expand Down Expand Up @@ -192,6 +260,44 @@ class CuSparseDnMatDescriptor {
cusparseDnMatDescr_t descriptor_;
};

/************* DENSE VECTOR DESCRIPTOR ************/
template <typename T>
class CuSparseDnVecDescriptor {
public:
explicit CuSparseDnVecDescriptor(const phi::DenseTensor& x,
const phi::GPUContext& dev_ctx)
: dev_ctx_(dev_ctx) {
std::vector<int64_t> xdim_vec = phi::vectorize(x.dims());
auto x_ndims = xdim_vec.size();
PADDLE_ENFORCE_GE(x_ndims,
1,
phi::errors::InvalidArgument(
"the dim size of Vec must be eaqual to 1."));

const T* x_data = x.data<T>();
cudaDataType_t gpu_type = GetGpuDataType<T>();
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseCreateDnVec(
&descriptor_, x.numel(), const_cast<T*>(x_data), gpu_type);
});

VLOG(6) << "Create cusparseDnVecDescr_t " << &descriptor_;
}

~CuSparseDnVecDescriptor() {
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseDestroyDnVec(descriptor_);
});
VLOG(6) << "Destroy cusparseDnVecDescr_t " << &descriptor_;
}

const cusparseDnVecDescr_t& descriptor() const { return descriptor_; }

private:
const phi::GPUContext& dev_ctx_;
cusparseDnVecDescr_t descriptor_;
};

template <>
template <typename T, typename TensorType>
void SparseBlas<phi::GPUContext>::SPMM(bool transa,
Expand Down Expand Up @@ -239,6 +345,50 @@ void SparseBlas<phi::GPUContext>::SPMM(bool transa,
});
}

template <>
template <typename T, typename TensorType>
void SparseBlas<phi::GPUContext>::SPMV(bool transa,
T alpha,
const TensorType& mat_a,
const phi::DenseTensor& vec_x,
T beta,
phi::DenseTensor* vec_out) const {
auto a_descriptor = CuSparseSpMatDescriptor<T>(mat_a, dev_ctx_);
auto x_descriptor = CuSparseDnVecDescriptor<T>(vec_x, dev_ctx_);
auto out_descriptor = CuSparseDnVecDescriptor<T>(*vec_out, dev_ctx_);

cudaDataType_t gpu_type = GetGpuDataType<T>();
size_t buffer_size = 0;
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSpMV_bufferSize(handle,
GetTransposeOperation(transa),
&alpha,
a_descriptor.descriptor(),
x_descriptor.descriptor(),
&beta,
out_descriptor.descriptor(),
gpu_type,
CUSPARSE_MV_ALG_DEFAULT,
&buffer_size);
});

paddle::memory::allocation::AllocationPtr tmp_buffer =
paddle::memory::Alloc(dev_ctx_, buffer_size);
void* tmp_buffer_ptr = tmp_buffer->ptr();
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSpMV(handle,
GetTransposeOperation(transa),
&alpha,
a_descriptor.descriptor(),
x_descriptor.descriptor(),
&beta,
out_descriptor.descriptor(),
gpu_type,
CUSPARSE_MV_ALG_DEFAULT,
tmp_buffer_ptr);
});
}

#if CUDA_VERSION >= 11030
template <>
template <typename T, typename TensorType>
Expand All @@ -249,12 +399,11 @@ void SparseBlas<phi::GPUContext>::SDDMM(bool transa,
const phi::DenseTensor& mat_b,
T beta,
TensorType* mat_out) const {
cudaDataType_t gpu_type = GetGpuDataType<T>();

auto a_descriptor = CuSparseDnMatDescriptor<T>(mat_a, dev_ctx_);
auto b_descriptor = CuSparseDnMatDescriptor<T>(mat_b, dev_ctx_);
auto out_descriptor = CuSparseSpMatDescriptor<T>(*mat_out, dev_ctx_);

cudaDataType_t gpu_type = GetGpuDataType<T>();
size_t buffer_size = 0;
dev_ctx_.CusparseCall([&](cusparseHandle_t handle) {
phi::dynload::cusparseSDDMM_bufferSize(handle,
Expand Down
56 changes: 56 additions & 0 deletions paddle/phi/kernels/sparse/cpu/mv_grad_kernel.cc
View file
Open in desktop
@@ -0,0 +1,56 @@
/* Copyright (c) 2022 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 "paddle/phi/kernels/sparse/mv_grad_kernel.h"

#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/core/kernel_registry.h"

namespace phi {
namespace sparse {

template <typename T, typename Context>
void MvCooGradKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& vec,
const DenseTensor& dout,
SparseCooTensor* dx,
DenseTensor* dvec) {
PADDLE_THROW(phi::errors::Unimplemented(
"Not support CPU backward kernel of 'sparse.mv' now."));
}

template <typename T, typename Context>
void MvCsrGradKernel(const Context& dev_ctx,
const SparseCsrTensor& x,
const DenseTensor& vec,
const DenseTensor& dout,
SparseCsrTensor* dx,
DenseTensor* dvec) {
PADDLE_THROW(phi::errors::Unimplemented(
"Not support CPU backward kernel of 'sparse.mv' now."));
}

} // namespace sparse
} // namespace phi

PD_REGISTER_KERNEL(
mv_coo_grad, CPU, ALL_LAYOUT, phi::sparse::MvCooGradKernel, float, double) {
kernel->InputAt(0).SetDataLayout(phi::DataLayout::SPARSE_COO);
}

PD_REGISTER_KERNEL(
mv_csr_grad, CPU, ALL_LAYOUT, phi::sparse::MvCsrGradKernel, float, double) {
kernel->InputAt(0).SetDataLayout(phi::DataLayout::SPARSE_CSR);
}