torch-linear-assignment 0.0.2__tar.gz → 0.0.4__tar.gz

{torch_linear_assignment-0.0.2/torch_linear_assignment.egg-info → torch_linear_assignment-0.0.4}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.4
 Name: torch-linear-assignment
-Version: 0.0.2
+Version: 0.0.4
 Summary: Batched linear assignment with PyTorch and CUDA.
 Author: Ivan Karpukhin
 Author-email: karpuhini@yandex.ru
@@ -8,6 +8,13 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: torch>=1.12.0
 Requires-Dist: scipy>=1.6.0
+Dynamic: author
+Dynamic: author-email
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: license-file
+Dynamic: requires-dist
+Dynamic: summary
 
 # Batch linear assignment for PyTorch
 [![PyPI version](https://badge.fury.io/py/torch-linear-assignment.svg)](https://badge.fury.io/py/torch-linear-assignment)
@@ -34,6 +41,11 @@ Build and install from Git repository:
 pip install .
 ```
 
+Building in an isolated environment may use a different PyTorch version. To match the current environment and reduce the disk usage, apply the following flag:
+```bash
+pip install --no-build-isolation torch-linear-assignment
+```
+
 When building with CUDA, make sure NVCC has the same CUDA version as PyTorch.
 You can choose CUDA version by
 ```bash

{torch_linear_assignment-0.0.2 → torch_linear_assignment-0.0.4}/README.md

@@ -23,6 +23,11 @@ Build and install from Git repository:
 pip install .
 ```
 
+Building in an isolated environment may use a different PyTorch version. To match the current environment and reduce the disk usage, apply the following flag:
+```bash
+pip install --no-build-isolation torch-linear-assignment
+```
+
 When building with CUDA, make sure NVCC has the same CUDA version as PyTorch.
 You can choose CUDA version by
 ```bash

{torch_linear_assignment-0.0.2 → torch_linear_assignment-0.0.4}/setup.py

@@ -48,7 +48,7 @@ with open("README.md") as fp:
 if __name__ == '__main__':
     setuptools.setup(
         name="torch-linear-assignment",
-        version="0.0.2",
+        version="0.0.4",
         author="Ivan Karpukhin",
         author_email="karpuhini@yandex.ru",
         description="Batched linear assignment with PyTorch and CUDA.",

{torch_linear_assignment-0.0.2 → torch_linear_assignment-0.0.4}/src/torch_linear_assignment_cuda_kernel.cu

@@ -9,8 +9,6 @@
 
 #include <cuda.h>
 #include <cuda_runtime.h>
-#include <thrust/device_vector.h>
-#include <thrust/fill.h>
 
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
@@ -37,17 +35,20 @@ int SMPCores(int device_index)
   case 6: // Pascal
     if ((devProp.minor == 1) || (devProp.minor == 2)) return 128;
     else if (devProp.minor == 0) return 64;
+    break;
   case 7: // Volta and Turing
     if ((devProp.minor == 0) || (devProp.minor == 5)) return 64;
+    break;
   case 8: // Ampere
     if (devProp.minor == 0) return 64;
     else if (devProp.minor == 6) return 128;
     else if (devProp.minor == 9) return 128; // ada lovelace
+    break;
   case 9: // Hopper
     if (devProp.minor == 0) return 128;
-  // Unknown device;
+    break;
   }
-  return 128;
+  return 128; // Unknown device
 }
 
 
@@ -203,8 +204,10 @@ void solve_cuda_kernel_batch(int bs, int nr, int nc,
 }
 
 
+
 template <typename scalar_t>
-void solve_cuda_batch(int device_index,
+void solve_cuda_batch(c10::ScalarType scalar_type,
+                      int device_index,
                       int bs, int nr, int nc,
                       scalar_t *cost, int *col4row, int *row4col) {
   cudaSetDevice(device_index);
@@ -212,32 +215,38 @@ void solve_cuda_batch(int device_index,
   TORCH_CHECK(std::numeric_limits<scalar_t>::has_infinity, "Data type doesn't have infinity.");
   auto infinity = std::numeric_limits<scalar_t>::infinity();
 
-  thrust::device_vector<scalar_t> u(bs * nr);
-  thrust::device_vector<scalar_t> v(bs * nc);
-  thrust::device_vector<scalar_t> shortestPathCosts(bs * nc);
-  thrust::device_vector<int> path(bs * nc);
-  thrust::device_vector<uint8_t> SR(bs * nr);
-  thrust::device_vector<uint8_t> SC(bs * nc);
-  thrust::device_vector<int> remaining(bs * nc);
-
-  thrust::fill(u.begin(), u.end(), (scalar_t) 0);
-  thrust::fill(v.begin(), v.end(), (scalar_t) 0);
-  thrust::fill(path.begin(), path.end(), -1);
-
-  int blockSize = SMPCores(device_index);
+  auto int_opt = torch::TensorOptions()
+    .dtype(torch::kInt)
+    .device(torch::kCUDA, device_index);
+  auto scalar_t_opt = torch::TensorOptions()
+    .dtype(scalar_type)
+    .device(torch::kCUDA, device_index);
+  auto uint8_opt = torch::TensorOptions()
+    .dtype(torch::kUInt8)
+    .device(torch::kCUDA, device_index);
+
+  torch::Tensor u = torch::zeros({bs * nr}, scalar_t_opt);
+  torch::Tensor v = torch::zeros({bs * nc}, scalar_t_opt);
+  torch::Tensor shortestPathCosts = torch::empty({bs * nc}, scalar_t_opt);
+  torch::Tensor path = torch::full({bs * nc}, -1, int_opt);
+  torch::Tensor SR = torch::empty({bs * nr}, uint8_opt);
+  torch::Tensor SC = torch::empty({bs * nc}, uint8_opt);
+  torch::Tensor remaining = torch::empty({bs * nc}, int_opt);
+
+  static const int blockSize = SMPCores(device_index);
   int gridSize = (bs + blockSize - 1) / blockSize;
   at::cuda::CUDAStream stream = at::cuda::getCurrentCUDAStream(device_index);
   solve_cuda_kernel_batch<<<gridSize, blockSize, 0, stream.stream()>>>(
     bs, nr, nc,
     cost,
-    thrust::raw_pointer_cast(&u.front()),
-    thrust::raw_pointer_cast(&v.front()),
-    thrust::raw_pointer_cast(&shortestPathCosts.front()),
-    thrust::raw_pointer_cast(&path.front()),
+    u.data<scalar_t>(),
+    v.data<scalar_t>(),
+    shortestPathCosts.data<scalar_t>(),
+    path.data<int>(),
     col4row, row4col,
-    thrust::raw_pointer_cast(&SR.front()),
-    thrust::raw_pointer_cast(&SC.front()),
-    thrust::raw_pointer_cast(&remaining.front()),
+    SR.data<uint8_t>(),
+    SC.data<uint8_t>(),
+    remaining.data<int>(),
     infinity);
   cudaError_t err = cudaGetLastError();
   if (err != cudaSuccess) {
@@ -265,6 +274,7 @@ std::vector<torch::Tensor> batch_linear_assignment_cuda(torch::Tensor cost) {
 
   AT_DISPATCH_FLOATING_TYPES(cost.scalar_type(), "solve_cuda_batch", [&] {
     solve_cuda_batch<scalar_t>(
+        cost.scalar_type(),
         device.index(),
         sizes[0], sizes[1], sizes[2],
         cost.data<scalar_t>(),

{torch_linear_assignment-0.0.2 → torch_linear_assignment-0.0.4/torch_linear_assignment.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.4
 Name: torch-linear-assignment
-Version: 0.0.2
+Version: 0.0.4
 Summary: Batched linear assignment with PyTorch and CUDA.
 Author: Ivan Karpukhin
 Author-email: karpuhini@yandex.ru
@@ -8,6 +8,13 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: torch>=1.12.0
 Requires-Dist: scipy>=1.6.0
+Dynamic: author
+Dynamic: author-email
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: license-file
+Dynamic: requires-dist
+Dynamic: summary
 
 # Batch linear assignment for PyTorch
 [![PyPI version](https://badge.fury.io/py/torch-linear-assignment.svg)](https://badge.fury.io/py/torch-linear-assignment)
@@ -34,6 +41,11 @@ Build and install from Git repository:
 pip install .
 ```
 
+Building in an isolated environment may use a different PyTorch version. To match the current environment and reduce the disk usage, apply the following flag:
+```bash
+pip install --no-build-isolation torch-linear-assignment
+```
+
 When building with CUDA, make sure NVCC has the same CUDA version as PyTorch.
 You can choose CUDA version by
 ```bash