Speed-up NMS by keeping index gathering on cuda device (#8766)

Co-authored-by: Nicolas Hug <[email protected]>
Co-authored-by: Nicolas Hug <[email protected]>

Author: Ghelfi

torchvision/csrc/ops/cuda/nms_kernel.cu

@@ -77,6 +77,51 @@ __global__ void nms_kernel_impl(
   }
 }
 
+__global__ static void gather_keep_from_mask(
+    bool* keep,
+    const unsigned long long* dev_mask,
+    const int n_boxes) {
+  // Taken and adapted from mmcv
+  // https://github.com/open-mmlab/mmcv/blob/03ce9208d18c0a63d7ffa087ea1c2f5661f2441a/mmcv/ops/csrc/common/cuda/nms_cuda_kernel.cuh#L76
+  const int col_blocks = ceil_div(n_boxes, threadsPerBlock);
+  const int thread_id = threadIdx.x;
+
+  // Mark the bboxes which have been removed.
+  extern __shared__ unsigned long long removed[];
+
+  // Initialize removed.
+  for (int i = thread_id; i < col_blocks; i += blockDim.x) {
+    removed[i] = 0;
+  }
+  __syncthreads();
+
+  for (int nblock = 0; nblock < col_blocks; nblock++) {
+    auto removed_val = removed[nblock];
+    __syncthreads();
+    const int i_offset = nblock * threadsPerBlock;
+#pragma unroll
+    for (int inblock = 0; inblock < threadsPerBlock; inblock++) {
+      const int i = i_offset + inblock;
+      if (i >= n_boxes)
+        break;
+      // Select a candidate, check if it should kept.
+      if (!(removed_val & (1ULL << inblock))) {
+        if (thread_id == 0) {
+          keep[i] = true;
+        }
+        auto p = dev_mask + i * col_blocks;
+        // Remove all bboxes which overlap the candidate.
+        for (int j = thread_id; j < col_blocks; j += blockDim.x) {
+          if (j >= nblock)
+            removed[j] |= p[j];
+        }
+        __syncthreads();
+        removed_val = removed[nblock];
+      }
+    }
+  }
+}
+
 at::Tensor nms_kernel(
     const at::Tensor& dets,
     const at::Tensor& scores,
@@ -133,35 +178,25 @@ at::Tensor nms_kernel(
             (unsigned long long*)mask.data_ptr<int64_t>());
       });
 
-  at::Tensor mask_cpu = mask.to(at::kCPU);
-  unsigned long long* mask_host =
-      (unsigned long long*)mask_cpu.data_ptr<int64_t>();
-
-  std::vector<unsigned long long> remv(col_blocks);
-  memset(&remv[0], 0, sizeof(unsigned long long) * col_blocks);
-
   at::Tensor keep =
-      at::empty({dets_num}, dets.options().dtype(at::kLong).device(at::kCPU));
-  int64_t* keep_out = keep.data_ptr<int64_t>();
-
-  int num_to_keep = 0;
-  for (int i = 0; i < dets_num; i++) {
-    int nblock = i / threadsPerBlock;
-    int inblock = i % threadsPerBlock;
-
-    if (!(remv[nblock] & (1ULL << inblock))) {
-      keep_out[num_to_keep++] = i;
-      unsigned long long* p = mask_host + i * col_blocks;
-      for (int j = nblock; j < col_blocks; j++) {
-        remv[j] |= p[j];
-      }
-    }
-  }
+      at::zeros({dets_num}, dets.options().dtype(at::kBool).device(at::kCUDA));
+
+  // Unwrap the mask to fill keep with proper values
+  // Keeping the unwrap on device instead of applying iterative for loops on cpu
+  // prevents the device -> cpu -> device transfer that could be bottleneck for
+  // large number of boxes.
+  // See https://github.com/pytorch/vision/issues/8713 for more details.
+  gather_keep_from_mask<<<
+      1,
+      min(col_blocks, threadsPerBlock),
+      col_blocks * sizeof(unsigned long long),
+      stream>>>(
+      keep.data_ptr<bool>(),
+      (unsigned long long*)mask.data_ptr<int64_t>(),
+      dets_num);
 
   AT_CUDA_CHECK(cudaGetLastError());
-  return order_t.index(
-      {keep.narrow(/*dim=*/0, /*start=*/0, /*length=*/num_to_keep)
-           .to(order_t.device(), keep.scalar_type())});
+  return order_t.masked_select(keep);
 }
 
 } // namespace