Revert "[Fix] Fix arf op's write conflict when num_orientations is no…

…t 1 (#2822)" (#2826)
open-mmlab · Jun 10, 2023 · 9c0f9cb · 9c0f9cb
1 parent e439825
commit 9c0f9cb
Show file tree

Hide file tree

Showing 3 changed files with 12 additions and 17 deletions.
diff --git a/mmcv/ops/csrc/common/cuda/active_rotated_filter_cuda_kernel.cuh b/mmcv/ops/csrc/common/cuda/active_rotated_filter_cuda_kernel.cuh
@@ -15,19 +15,18 @@ __global__ void active_rotated_filter_forward_cuda_kernel(
     const int nthreads, const scalar_t* weight_data, const int* indices_data,
     const int num_input_planes, const int num_output_planes,
     const int num_orientations, const int num_rotations, const int nEntry,
-    const int kH, const int kW, scalar_t* output_data) {
+    scalar_t* output_data) {
   CUDA_1D_KERNEL_LOOP(index, nthreads) {
     int l = index % nEntry;
     int j = (index / nEntry) % num_input_planes;
     int i = index / nEntry / num_input_planes;
     int k;
-    int fmIndex = (l / (kH * kW)) * kH * kW;
     scalar_t val = *(weight_data + index);
     for (k = 0; k < num_rotations; k++) {
       int idx = (int)(*(indices_data + l * num_rotations + k)) - 1;
-      scalar_t* target =
-          output_data + i * (num_rotations * num_input_planes * nEntry) +
-          k * (num_input_planes * nEntry) + j * (nEntry) + idx + fmIndex;
+      scalar_t* target = output_data +
+                         i * (num_rotations * num_input_planes * nEntry) +
+                         k * (num_input_planes * nEntry) + j * (nEntry) + idx;
       *target = val;
     }
   }
@@ -38,22 +37,20 @@ __global__ void active_rotated_filter_backward_cuda_kernel(
     const int nthreads, const scalar_t* gradWeight_data,
     const int* indices_data, const int num_input_planes,
     const int num_output_planes, const int num_orientations,
-    const int num_rotations, const int nEntry, const int kH, const int kW,
-    scalar_t* weight_data) {
+    const int num_rotations, const int nEntry, scalar_t* weight_data) {
   CUDA_1D_KERNEL_LOOP(index, nthreads) {
     int l = index % nEntry;
     int j = (index / nEntry) % num_input_planes;
     int i = index / nEntry / num_input_planes;
     int k;
-    int fmIndex = (l / (kH * kW)) * kH * kW;
     scalar_t* val = weight_data + index;
     *val = 0;
     scalar_t tmp = 0;
     for (k = 0; k < num_rotations; k++) {
       int idx = (int)(*(indices_data + l * num_rotations + k)) - 1;
       scalar_t target =
           *(gradWeight_data + i * (num_rotations * num_input_planes * nEntry) +
-            k * (num_input_planes * nEntry) + j * (nEntry) + idx + fmIndex);
+            k * (num_input_planes * nEntry) + j * (nEntry) + idx);
       tmp = tmp + target;
     }
     *val = tmp;

diff --git a/mmcv/ops/csrc/pytorch/cpu/active_rotated_filter.cpp b/mmcv/ops/csrc/pytorch/cpu/active_rotated_filter.cpp
@@ -19,12 +19,11 @@ void active_rotated_filter_forward_cpu_kernel(
       for (l = 0; l < nEntry; l++) {
         int weightIndex = i * num_input_planes * nEntry + j * nEntry + l;
         T val = *(weightData + weightIndex);
-        int fmIndex = (l / (kH * kW)) * kH * kW;
         for (k = 0; k < num_rotations; k++) {
           int index = (int)(*(indicesData + l * num_rotations + k)) - 1;
-          T* target =
-              outputData + i * (num_rotations * num_input_planes * nEntry) +
-              k * (num_input_planes * nEntry) + j * (nEntry) + index + fmIndex;
+          T* target = outputData +
+                      i * (num_rotations * num_input_planes * nEntry) +
+                      k * (num_input_planes * nEntry) + j * (nEntry) + index;
           *target = val;
         }
       }
@@ -49,12 +48,11 @@ void active_rotated_filter_backward_cpu_kernel(
         int gradInputIndex = i * num_input_planes * nEntry + j * nEntry + l;
         T* val = gradInputData + gradInputIndex;
         *val = 0;
-        int fmIndex = (l / (kH * kW)) * kH * kW;
         for (k = 0; k < num_rotations; k++) {
           int index = (int)(*(indicesData + l * num_rotations + k)) - 1;
           const T* target =
               gradOutputData + i * (num_rotations * num_input_planes * nEntry) +
-              k * (num_input_planes * nEntry) + j * (nEntry) + index + fmIndex;
+              k * (num_input_planes * nEntry) + j * (nEntry) + index;
           *val = *val + *target;
         }
       }

diff --git a/mmcv/ops/csrc/pytorch/cuda/active_rotated_filter_cuda.cu b/mmcv/ops/csrc/pytorch/cuda/active_rotated_filter_cuda.cu
@@ -24,7 +24,7 @@ void ActiveRotatedFilterForwardCUDAKernelLauncher(const Tensor input,
             <<<GET_BLOCKS(output_size), THREADS_PER_BLOCK, 0, stream>>>(
                 output_size, input.data_ptr<scalar_t>(),
                 indices.data_ptr<int>(), num_input_planes, num_output_planes,
-                num_orientations, num_rotations, nEntry, kH, kW,
+                num_orientations, num_rotations, nEntry,
                 output.data_ptr<scalar_t>());
       });
   AT_CUDA_CHECK(cudaGetLastError());
@@ -51,7 +51,7 @@ void ActiveRotatedFilterBackwardCUDAKernelLauncher(const Tensor grad_out,
             <<<GET_BLOCKS(output_size), THREADS_PER_BLOCK, 0, stream>>>(
                 output_size, grad_out.data_ptr<scalar_t>(),
                 indices.data_ptr<int>(), num_input_planes, num_output_planes,
-                num_orientations, num_rotations, nEntry, kH, kW,
+                num_orientations, num_rotations, nEntry,
                 grad_in.data_ptr<scalar_t>());
       });
   AT_CUDA_CHECK(cudaGetLastError());