Skip to content

P
Paddle

兽拳 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

未验证 提交 9e307229 编写于 作者: Z zhangkaihuo 提交者： GitHub
浏览文件
操作

Standard name of sparse pool (#44344)

上级 f382eb06
隐藏空白更改
内联 并排
Showing with 152 addition and 152 deletion
paddle/phi/api/yaml/sparse_api.yaml
浏览文件 @ 9e307229
......@@ -316,10 +316,10 @@
args : (Tensor x, int[] kernel_sizes, int[] paddings, int[] dilations, int[] strides)
output : Tensor(out), Tensor(rulebook)
kernel :
func : sparse_maxpool{sparse_coo -> sparse_coo, dense}
func : maxpool_coo{sparse_coo -> sparse_coo, dense}
layout : x
intermediate : rulebook
backward : sparse_maxpool_grad
backward : maxpool_grad
- api: mv
args : (Tensor x, Tensor vec)
......
paddle/phi/api/yaml/sparse_bw_api.yaml
浏览文件 @ 9e307229
......@@ -137,6 +137,13 @@
matmul_coo_dense_grad {sparse_coo, dense, dense -> sparse_coo, dense},
matmul_coo_coo_grad {sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo}
- backward_api : maxpool_grad
forward : maxpool(Tensor x, int[] kernel_sizes, int[] paddings, int[] dilations, int[] strides) -> Tensor(out), Tensor(rulebook)
args : (Tensor x, Tensor rulebook, Tensor out, Tensor out_grad, int[] kernel_sizes)
output : Tensor(x_grad)
kernel :
func : maxpool_coo_grad {sparse_coo, dense, sparse_coo, sparse_coo -> sparse_coo}
- backward_api : multiply_grad
forward : multiply(Tensor x, Tensor y) -> Tensor(out)
args : (Tensor x, Tensor y, Tensor out_grad)
......@@ -198,13 +205,6 @@
kernel :
func : softmax_csr_grad{sparse_csr, sparse_csr -> sparse_csr}
- backward_api : sparse_maxpool_grad
forward : sparse_maxpool(Tensor x, int[] kernel_sizes, int[] paddings, int[] dilations, int[] strides) -> Tensor(out), Tensor(rulebook)
args : (Tensor x, Tensor rulebook, Tensor out, Tensor out_grad, int[] kernel_sizes)
output : Tensor(x_grad)
kernel :
func : sparse_maxpool_grad {sparse_coo, dense, sparse_coo, sparse_coo -> sparse_coo}
- backward_api : sqrt_grad
forward : sqrt(Tensor x) -> Tensor(out)
args : (Tensor out, Tensor out_grad)
......@@ -255,7 +255,7 @@
- backward_api: fused_attention_grad
forward : fused_attention_csr(Tensor query, Tensor key, Tensor value, Tensor sparse_mask, Tensor key_padding_mask, Tensor attn_mask) -> Tensor(out), Tensor(softmax)
args: (Tensor query, Tensor key, Tensor value, Tensor softmax, Tensor out_grad)
output : Tensor(query_grad), Tensor(key_grad), Tensor(value_grad)
output : Tensor(query_grad), Tensor(key_grad), Tensor(value_grad)
kernel :
func : fused_attention_csr_grad{dense, dense, dense, sparse_csr, dense -> dense, dense, dense}
layout : softmax
......
paddle/phi/kernels/sparse/cpu/sparse_pool_grad_kernel.cc → paddle/phi/kernels/sparse/cpu/pool_grad_kernel.cc
浏览文件 @ 9e307229
......@@ -12,7 +12,7 @@ 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/sparse_pool_grad_kernel.h"
#include "paddle/phi/kernels/sparse/pool_grad_kernel.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/core/tensor_utils.h"
......@@ -25,13 +25,13 @@ namespace phi {
namespace sparse {
template <typename T, typename IntT = int>
void MaxPoolGradCPUKernel(const CPUContext& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad) {
void MaxPoolCooGradCPUKernel(const CPUContext& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad) {
int kernel_size = kernel_sizes[0] * kernel_sizes[1] * kernel_sizes[2];
const int channels = x.dims()[4];
int rulebook_len = rulebook.dims()[1];
......@@ -75,16 +75,16 @@ void MaxPoolGradCPUKernel(const CPUContext& dev_ctx,
}
template <typename T, typename Context>
void MaxPoolGradKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad) {
void MaxPoolCooGradKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad) {
PD_VISIT_INTEGRAL_TYPES(
x.non_zero_indices().dtype(), "MaxPoolGradCPUKernel", ([&] {
MaxPoolGradCPUKernel<T, data_t>(
x.non_zero_indices().dtype(), "MaxPoolCooGradCPUKernel", ([&] {
MaxPoolCooGradCPUKernel<T, data_t>(
dev_ctx, x, rulebook, out, out_grad, kernel_sizes, x_grad);
}));
}
......@@ -92,10 +92,10 @@ void MaxPoolGradKernel(const Context& dev_ctx,
} // namespace sparse
} // namespace phi
PD_REGISTER_KERNEL(sparse_maxpool_grad,
PD_REGISTER_KERNEL(maxpool_coo_grad,
CPU,
ALL_LAYOUT,
phi::sparse::MaxPoolGradKernel,
phi::sparse::MaxPoolCooGradKernel,
float,
double) {
kernel->InputAt(0).SetDataLayout(phi::DataLayout::SPARSE_COO);
......
paddle/phi/kernels/sparse/cpu/sparse_pool_kernel.cc → paddle/phi/kernels/sparse/cpu/pool_kernel.cc
浏览文件 @ 9e307229
......@@ -12,7 +12,7 @@ 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/sparse_pool_kernel.h"
#include "paddle/phi/kernels/sparse/pool_kernel.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/core/tensor_meta.h"
......@@ -30,14 +30,14 @@ namespace sparse {
* out: (N, D, H, W, OC)
**/
template <typename T, typename IntT = int>
void MaxPoolCPUKernel(const CPUContext& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook) {
void MaxPoolCooCPUKernel(const CPUContext& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook) {
const auto& x_dims = x.dims();
int kernel_size = kernel_sizes[0] * kernel_sizes[1] * kernel_sizes[2];
const std::vector<int>& real_kernel_sizes =
......@@ -98,34 +98,34 @@ void MaxPoolCPUKernel(const CPUContext& dev_ctx,
}
template <typename T, typename Context>
void MaxPoolKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook) {
void MaxPoolCooKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook) {
PD_VISIT_INTEGRAL_TYPES(
x.non_zero_indices().dtype(), "MaxPoolCPUKernel", ([&] {
MaxPoolCPUKernel<T, data_t>(dev_ctx,
x,
kernel_sizes,
paddings,
dilations,
strides,
out,
rulebook);
x.non_zero_indices().dtype(), "MaxPoolCooCPUKernel", ([&] {
MaxPoolCooCPUKernel<T, data_t>(dev_ctx,
x,
kernel_sizes,
paddings,
dilations,
strides,
out,
rulebook);
}));
}
} // namespace sparse
} // namespace phi
PD_REGISTER_KERNEL(sparse_maxpool,
PD_REGISTER_KERNEL(maxpool_coo,
CPU,
ALL_LAYOUT,
phi::sparse::MaxPoolKernel,
phi::sparse::MaxPoolCooKernel,
float,
double) {
kernel->InputAt(0).SetDataLayout(phi::DataLayout::SPARSE_COO);
......
paddle/phi/kernels/sparse/gpu/sparse_pool_grad_kernel.cu → paddle/phi/kernels/sparse/gpu/pool_grad_kernel.cu
浏览文件 @ 9e307229
......@@ -12,7 +12,7 @@ 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/sparse_pool_grad_kernel.h"
#include "paddle/phi/kernels/sparse/pool_grad_kernel.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/backends/gpu/gpu_info.h"
......@@ -52,13 +52,13 @@ __global__ void MaxPoolGradCudaKernel(const T* in_features_ptr,
}
template <typename T, typename IntT = int>
void MaxPoolGradGPUKernel(const GPUContext& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad) {
void MaxPoolCooGradGPUKernel(const GPUContext& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad) {
int kernel_size = kernel_sizes[0] * kernel_sizes[1] * kernel_sizes[2];
const int in_channels = x.dims()[4];
int rulebook_len = rulebook.dims()[1];
......@@ -121,16 +121,16 @@ void MaxPoolGradGPUKernel(const GPUContext& dev_ctx,
}
template <typename T, typename Context>
void MaxPoolGradKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad) {
void MaxPoolCooGradKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad) {
PD_VISIT_INTEGRAL_TYPES(
x.non_zero_indices().dtype(), "MaxPoolGradGPUKernel", ([&] {
MaxPoolGradGPUKernel<T, data_t>(
x.non_zero_indices().dtype(), "MaxPoolCooGradGPUKernel", ([&] {
MaxPoolCooGradGPUKernel<T, data_t>(
dev_ctx, x, rulebook, out, out_grad, kernel_sizes, x_grad);
}));
}
......@@ -138,10 +138,10 @@ void MaxPoolGradKernel(const Context& dev_ctx,
} // namespace sparse
} // namespace phi
PD_REGISTER_KERNEL(sparse_maxpool_grad,
PD_REGISTER_KERNEL(maxpool_coo_grad,
GPU,
ALL_LAYOUT,
phi::sparse::MaxPoolGradKernel,
phi::sparse::MaxPoolCooGradKernel,
float,
double) {
kernel->InputAt(0).SetDataLayout(phi::DataLayout::SPARSE_COO);
......
paddle/phi/kernels/sparse/gpu/sparse_pool_kernel.cu → paddle/phi/kernels/sparse/gpu/pool_kernel.cu
浏览文件 @ 9e307229
......@@ -12,7 +12,7 @@ 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/sparse_pool_kernel.h"
#include "paddle/phi/kernels/sparse/pool_kernel.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/core/tensor_meta.h"
......@@ -48,14 +48,14 @@ __global__ void MaxPoolCudaKernel(const T* in_features_ptr,
* out: (N, D, H, W, OC)
**/
template <typename T, typename IntT = int>
void MaxPoolGPUKernel(const GPUContext& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook) {
void MaxPoolCooGPUKernel(const GPUContext& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook) {
const auto& x_dims = x.dims();
int kernel_size = kernel_sizes[0] * kernel_sizes[1] * kernel_sizes[2];
const std::vector<int>& real_kernel_sizes =
......@@ -127,34 +127,34 @@ void MaxPoolGPUKernel(const GPUContext& dev_ctx,
}
template <typename T, typename Context>
void MaxPoolKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook) {
void MaxPoolCooKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook) {
PD_VISIT_INTEGRAL_TYPES(
x.non_zero_indices().dtype(), "MaxPoolGPUKernel", ([&] {
MaxPoolGPUKernel<T, data_t>(dev_ctx,
x,
kernel_sizes,
paddings,
dilations,
strides,
out,
rulebook);
x.non_zero_indices().dtype(), "MaxPoolCooGPUKernel", ([&] {
MaxPoolCooGPUKernel<T, data_t>(dev_ctx,
x,
kernel_sizes,
paddings,
dilations,
strides,
out,
rulebook);
}));
}
} // namespace sparse
} // namespace phi
PD_REGISTER_KERNEL(sparse_maxpool,
PD_REGISTER_KERNEL(maxpool_coo,
GPU,
ALL_LAYOUT,
phi::sparse::MaxPoolKernel,
phi::sparse::MaxPoolCooKernel,
float,
double,
phi::dtype::float16) {
......
paddle/phi/kernels/sparse/sparse_pool_grad_kernel.h → paddle/phi/kernels/sparse/pool_grad_kernel.h
浏览文件 @ 9e307229
......@@ -22,23 +22,23 @@ namespace phi {
namespace sparse {
template <typename T, typename Context>
void MaxPoolGradKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad);
void MaxPoolCooGradKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes,
SparseCooTensor* x_grad);
template <typename T, typename Context>
SparseCooTensor MaxPoolGrad(const Context& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes) {
SparseCooTensor MaxPoolCooGrad(const Context& dev_ctx,
const SparseCooTensor& x,
const DenseTensor& rulebook,
const SparseCooTensor& out,
const SparseCooTensor& out_grad,
const std::vector<int>& kernel_sizes) {
SparseCooTensor x_grad;
MaxPoolGradKernel<T, Context>(
MaxPoolCooGradKernel<T, Context>(
dev_ctx, x, rulebook, out, out_grad, kernel_sizes, &x_grad);
return x_grad;
}
......
paddle/phi/kernels/sparse/sparse_pool_kernel.h → paddle/phi/kernels/sparse/pool_kernel.h
浏览文件 @ 9e307229
......@@ -22,25 +22,25 @@ namespace phi {
namespace sparse {
template <typename T, typename Context>
void MaxPoolKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook);
void MaxPoolCooKernel(const Context& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
SparseCooTensor* out,
DenseTensor* rulebook);
template <typename T, typename Context>
SparseCooTensor MaxPool(const Context& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
DenseTensor* rulebook) {
SparseCooTensor MaxPoolCoo(const Context& dev_ctx,
const SparseCooTensor& x,
const std::vector<int>& kernel_sizes,
const std::vector<int>& paddings,
const std::vector<int>& dilations,
const std::vector<int>& strides,
DenseTensor* rulebook) {
SparseCooTensor coo;
MaxPoolKernel<T, Context>(
MaxPoolCooKernel<T, Context>(
dev_ctx, x, kernel_sizes, paddings, dilations, strides, &coo, rulebook);
return coo;
}
......
paddle/phi/tests/kernels/test_sparse_pool_dev_api.cc
浏览文件 @ 9e307229
......@@ -23,8 +23,8 @@ limitations under the License. */
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/core/tensor_utils.h"
#include "paddle/phi/kernels/sparse/coalesce_kernel.h"
#include "paddle/phi/kernels/sparse/sparse_pool_grad_kernel.h"
#include "paddle/phi/kernels/sparse/sparse_pool_kernel.h"
#include "paddle/phi/kernels/sparse/pool_grad_kernel.h"
#include "paddle/phi/kernels/sparse/pool_kernel.h"
namespace phi {
namespace tests {
......@@ -91,13 +91,13 @@ void TestMaxPoolBase(const std::vector<IntT>& indices,
if (!std::is_same<T, phi::dtype::float16>::value) {
DenseTensor rulebook;
SparseCooTensor out = sparse::MaxPool<T>(dev_ctx_cpu,
x_tensor,
kernel_sizes,
paddings,
dilations,
strides,
&rulebook);
SparseCooTensor out = sparse::MaxPoolCoo<T>(dev_ctx_cpu,
x_tensor,
kernel_sizes,
paddings,
dilations,
strides,
&rulebook);
ASSERT_EQ(correct_out_dims.size(), out.dims().size());
for (int i = 0; i < correct_out_dims.size(); i++) {
......@@ -113,7 +113,7 @@ void TestMaxPoolBase(const std::vector<IntT>& indices,
f_verify(out.non_zero_elements().data<T>(), correct_out_features);
if (backward) {
SparseCooTensor x_grad = sparse::MaxPoolGrad<T>(
SparseCooTensor x_grad = sparse::MaxPoolCooGrad<T>(
dev_ctx_cpu, x_tensor, rulebook, out, out, kernel_sizes);
f_verify(x_grad.non_zero_elements().data<T>(), features_grad);
}
......@@ -151,13 +151,13 @@ void TestMaxPoolBase(const std::vector<IntT>& indices,
SparseCooTensor d_x_tensor(d_indices_tensor, d_features_tensor, x_dims);
DenseTensor d_rulebook;
SparseCooTensor d_out = sparse::MaxPool<T>(dev_ctx_gpu,
d_x_tensor,
kernel_sizes,
paddings,
dilations,
strides,
&d_rulebook);
SparseCooTensor d_out = sparse::MaxPoolCoo<T>(dev_ctx_gpu,
d_x_tensor,
kernel_sizes,
paddings,
dilations,
strides,
&d_rulebook);
SparseCooTensor tmp_d_out = sparse::Coalesce<T>(dev_ctx_gpu, d_out);
ASSERT_EQ(correct_out_dims.size(), d_out.dims().size());
......@@ -191,7 +191,7 @@ void TestMaxPoolBase(const std::vector<IntT>& indices,
f_verify(h_features_tensor.data<T>(), correct_out_features);
if (backward) {
SparseCooTensor x_grad = sparse::MaxPoolGrad<T>(
SparseCooTensor x_grad = sparse::MaxPoolCooGrad<T>(
dev_ctx_gpu, d_x_tensor, d_rulebook, d_out, d_out, kernel_sizes);
DenseTensor h_features_grad =
phi::EmptyLike<T>(dev_ctx_cpu, x_grad.non_zero_elements());
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或